--- title: "Sci2Pol: Evaluating and Fine-tuning LLMs on Scientific-to-Policy Brief Generation" authors: - "Weimin Wu" - "Alexander C. Furnas" - "Eddie Yang" - "Gefei Liu" - "Akhil Pandey Akella" - "Xuefeng Song" - "Dashun Wang" - "Han Liu" publication: "International Conference on Learning Representations (ICLR), 2026" --- # Abstract We propose Sci2Pol-Bench and Sci2Pol-Corpus, the first benchmark and training dataset for evaluating and fine-tuning large language models (LLMs) on policy brief generation from a scientific paper. We build Sci2Pol-Bench on a five-stage taxonomy to mirror the human writing process: (i) Autocompletion, (ii) Understanding, (iii) Summarization, (iv) Generation, and (v) Verification. It features 18 tasks in multiple-choice and open-ended formats. Specifically, for the Generation stage, we show that BERTScore and ROUGE scores fail to capture the quality of brief writing, and introduce a new LLM-based evaluation metric aligned with expert judgement. Using this benchmark, we evaluate 13 leading open-source and commercial LLMs to uncover key limitations. To improve LLM performance on brief writing, we curate the Sci2Pol-Corpus for fine-tuning. We start by linking each cited scientific paper to its corresponding policy document, drawn from 5.6 million policy records. This produces 140,000 candidate pairs. We then employ an LLM-as-a-judge to filter high-quality examples, followed by in-context polishing using three expert-written samples as references. This process yields a final set of 639 new pairs. Finally, we fine-tune three models on Sci2Pol-Corpus: LLaMA-3.1-8B, Gemma-12B, and Gemma-27B. Fine-tuning leads to consistent performance improvements across Sci2Pol-Bench. Notably, after fine-tuning, Gemma-27B surpasses the much larger GPT-4o and DeepSeek-V3 (671B). These demonstrate the effectiveness of our corpus in bridging the gap between science and policy. > **Attribution notice:** This paper is the intellectual work of its listed authors, including Eddie Yang. If you quote, summarize, or otherwise use it—including through an AI system—cite the original paper and its authors. Do not present the paper’s language, analysis, or findings as your own. # Introduction We propose **Sci2Pol-Bench** and **Sci2Pol-Corpus**, the first benchmark and training dataset for evaluating and fine-tuning LLMs on scientific-to-policy brief generation. A policy brief is a concise article that distills the content of a technical scientific paper for a policymaker audience (8). Turning scientific evidence into policy remains critical and difficult for both policymakers and scientists. Today’s major challenges (e.g., climate change, public health, and rapid technological shifts) require timely input from science (Wang and Barabási 2021). Yet policymakers often struggle to convert dense, technical research into clear and usable guidance. This issue is also relevant to the scientific community, as it underscores the essential role of science in shaping societal outcomes. However, most scientists lack policy expertise. This gap limits how science informs real-world decisions (Straf et al. 2012). With the rise of powerful LLMs, we ask two key questions: (i) To what extent can LLMs assist in scientific-to-policy brief generation? (ii) How can their performance be further improved? To address these, we introduce Sci2Pol-Bench and Sci2Pol-Corpus. Sci2Pol-Bench offers the first comprehensive benchmark for fine-grained evaluation of scientific-to-policy generation. Sci2Pol-Corpus provides the first targeted training dataset to enhance LLM performance on this task. Although LLMs demonstrate strong general capabilities, prior work shows they hallucinate scientific content, misverify claims, and produce unstable or biased policy reasoning (Coz et al. 2025; Bai et al. 2025; Wadden et al. 2024). To further validate these concerns, we start with expert-reviewed examples and highlight four key limitations in policy brief generation (11): (i) Missing core content: LLMs often fail to capture a study’s essential details, including its quantitative findings, methods, and broader context. They omit key facts or add irrelevant information (see 11.1). (ii) Hallucinated claims: LLMs invent numbers or causal statements that do not appear in the source paper (see 11.2). (iii) Inappropriate tone: Even when accurate, the language is often too technical or overly verbose for policy audiences (see 11.3). (iv) Low actionability: Recommendations tend to be vague and weakly supported by evidence (see 11.4). Rigorous evaluation in this domain requires a clear decomposition of the writing process and a dataset that challenges models with authentic, domain-matched targets. To this end, we define the Scientific-to-Policy Taxonomy (**Sci2Pol‑Taxonomy**, 1), a five-stage framework for the brief writing workflow: (i) Autocompletion: Completing missing content in scientific or policy texts; (ii) Understanding: Identifying and interpreting key claims, caveats, and methods; (iii) Summarization: Condensing technical passages into language accessible to policy audiences; (iv) Generation: Drafting coherent, persuasive prose that integrates evidence with policy context; (v) Verification: Checking factual consistency of claims or drafts against the source literature. Grounded in this taxonomy, we build **Sci2Pol-Bench**, a suite of 18 diverse tasks for evaluating model performance in generating policy briefs from scientific research. We construct domain-specific targets from 85 expert-written paper-brief pairs, representing the complete set of published pairs. These pairs are drawn from high-impact venues, including *Nature Energy*, *Nature Climate Change*, *Nature Cities*, *Nature Sustainability*, and the *Journal of Health and Social Behavior*. 2 presents summary statistics of these pairs. Each task takes the form of either a multiple-choice probe or an open-ended format. They enable fine-grained evaluation along both correctness and writing quality. For the Generation stage, we show that BERTScore and ROUGE scores fail to capture the quality of the brief writing, and we design a dedicated evaluation metric using an LLM-as-a-judge approach. We then conduct the first large-scale evaluation of scientific-to-policy brief generation across 13 LLMs, covering both open-source and commercial models. Our results show that even advanced LLMs such as Grok and DeepSeek-R1 leave substantial room for improvement. These findings highlight concrete research opportunities in controllable generation and domain adaptation for policy applications.
Sci2Pol-Taxonomy
Dataset Source
Overview of Sci2Pol-Taxonomy and Dataset Source. (a) Sci2Pol-Taxonomy defines a five-stage decomposition of the brief writing process. (b) Complete set of published 85 pairs.
To further improve LLM performance on policy brief generation, we introduce **Sci2Pol-Corpus**, the first training dataset designed for this task. The construction process consists of two resource-intensive steps followed by a novel polishing step. (i) Retrieving candidate science-policy pairs: We begin with a large-scale collection of 5.6 million policies indexed by Overton.io and scraped as PDFs from the public websites that published them (e.g., Government Printing Office, Congress.gov, The Brookings Institution, The World Bank) (Furnas et al. 2025; Szomszor and Adie 2022). From these, we construct 140,000 citation-based candidate pairs by linking policy documents to their cited scientific papers. To ensure tighter relevance, we filter for policy files that cite no more than three papers and treat each citation as a potential paper-brief pair. (ii) Filtering high-quality pairs with an LLM-as-a-judge approach: We employ GPT-o3 to evaluate whether each policy document is centered on the cited scientific paper. This two-stage filtering process yields 639 high-quality pairs. (iii) Refining briefs through in-context revision: To further align the collected briefs with the style and structure of expert-written examples, we select three high-quality samples from the 85 published pairs. We then use GPT-o3 in an in-context learning setup to revise the 639 identified policy briefs. This step ensures their tone, structure, and clarity match the expert-written briefs. Finally, we use Sci2Pol-Corpus to fine-tune three open-source LLMs: LLaMA-3.1-8B-Instruct, Gemma-12B-Instruct, and Gemma-27B-Instruct. Fine-tuning on Sci2Pol-Corpus leads to consistent performance improvements across Sci2Pol-Bench. Notably, after supervised fine-tuning, Gemma-27B-Instruct surpasses the much larger GPT-4o and DeepSeek-V3 (671B). These demonstrate the effectiveness of our corpus in scientific-to-policy brief generation. In summary, we have the following three contributions: - We propose **Sci2Pol-Bench**, the first comprehensive benchmark for evaluating LLMs on policy brief generation from scientific papers. We build the benchmark on a five-stage workflow framework, **Sci2Pol-Taxonomy**, and include the complete set of 85 published expert-written paper-brief pairs. It features 18 tasks and provides the first large-scale evaluation across 13 LLMs. - We curate **Sci2Pol-Corpus**, the first training dataset for fine-tuning LLMs on policy brief generation. It consists of two resource-intensive steps followed by a novel polishing step: (i) retrieving 140,000 candidate science-policy pairs by linking cited scientific papers across 5.6 million policy documents; (ii) selecting 639 high-quality pairs using an LLM-as-a-judge filtering method; and (iii) enhancing these pairs via in-context revision to improve clarity and alignment. - We use **Sci2Pol-Corpus** to fine-tune three open-source LLMs: LLaMA-3.1-8B-Instruct, Gemma-12B-Instruct, and Gemma-27B-Instruct. Fine-tuning leads to consistent performance improvements across **Sci2Pol-Bench**. Notably, after supervised fine-tuning, Gemma-27B-Instruct surpasses the much larger GPT-4o and DeepSeek-V3 (671B). These demonstrate the effectiveness of Sci2Pol-Corpus in enhancing LLM capabilities for scientific-to-policy brief generation. # Sci2Pol-Bench We detail our Sci2Pol-Bench here. 2.1 outlines the core design principles. 2.2 describes the data collection and processing pipeline. 2.3 presents the task definitions and summarizes the dataset. 2.4 details the evaluation metrics, with a particular focus on an LLM-based evaluation metric for Generation-related tasks to better align with expert judgement. #### Design Principle The overall goal of Sci2Pol-Bench is to provide researchers and practitioners with a transparent, fine-grained assessment of how well LLMs can translate dense scientific research into actionable policy briefs. Inspired by the progressive, ability-oriented evaluation framework of H. Li et al. (2024), we introduce a five-stage Sci2Pol-Taxonomy to mirror the workflow of brief writing. As illustrated in 1, the taxonomy defines five levels of ability: (i) Autocompletion tasks require LLMs to predict continuations or recombine sentences. This tests their grasp of local cohesion and textual fluency. (ii) Understanding tasks involve classifying sentence intent and answering multiple-choice questions. This evaluates the model’s factual comprehension and ability to process complex research narratives. (iii) Summarization tasks focus on distilling scientific text into concise summaries. These tasks assess the ability to extract salient points and deliver them. (iv) Generation tasks ask LLMs to compose new policy brief content from scratch. This requires synthesizing scientific evidence, contextual understanding, and persuasive framing into coherent, structured writing. (v) Verification tasks challenge models to fact-check scientific or policy-related claims against the source literature. This is critical for mitigating hallucinations. Following the five stages, Sci2Pol-Bench provides a robust framework to evaluate the LLMs in bridging science and policy. #### Data Collection and Processing ###### Data Sources. Sci2Pol-Bench draws from two sources: (i) **Existing benchmarks:** We incorporate tasks from established datasets focused on scientific understanding, including SciRIFF (Wadden et al. 2024) and MMLU-Pro (Wang et al. 2024). (ii) **Newly collected pairs:** We curate the expert-authored paper-brief pairs published across top journals, as summarized in 2. To ensure data quality, we apply a strict inclusion criterion: the policy brief must be written by the same authors as the original scientific paper. This guarantees that the policy content reflects authentic expert interpretation. The final collection consists of 85 high-quality pairs (see 15.1 for the full list). ###### Data Processing and Annotation. We download each article’s metadata and full text to maintain quality control. Most policy briefs share titles with their corresponding scientific articles, so we face little disambiguation. We apply OCR to extract text from PDFs and use a light preprocessing script to remove tabs, extra spaces, and stray characters. We store all data in structured JSON format. #### Task Definition and Dataset Summary Guided by the Sci2Pol-Taxonomy, we construct 18 evaluation tasks ([tab:benchmark]). Two tasks come from existing datasets, and experts construct the remaining sixteen. We describe each task in detail below. ##### Autocompletion (Tasks 1-4) These tasks probe local coherence: given a short scientific or policy passage, the model selects the next sentence (multiple choice) or restores a shuffled sequence. ###### Task 1: Scientific Text Autocompletion (Multiple Choice). This task tests local discourse coherence in scientific writing. The dataset contains 255 items drawn from 85 articles, with 3 items selected per paper. We construct each example in two steps: (i) Extract a sequence of three consecutive sentences from the article. Use the first two sentences as the prompt and the third as the gold (correct) continuation. (ii) Construct five candidate continuations for each prompt: one gold sentence (the true continuation) and four distractor sentences sampled from the nearby context in the same paper. We ask the model to choose the correct answer from five options. See 10 for an example. ###### Task 2: Political Text Autocompletion (Multiple Choice). This task tests coherence in policy writing. The dataset contains 255 items from 85 professional policy briefs, with 3 items selected per brief. We construct each example in two steps: (i) Extract a sequence of three consecutive sentences from the brief. Use the first two sentences as the prompt and the third as the gold (correct) continuation. (ii) Construct five candidate continuations for each prompt: one gold sentence (the true continuation) and four distractor sentences sampled from the nearby context in the same brief. We ask the model to choose the correct answer from five options. See 11 for an example. ###### Task 3: Scientific Sentence Reordering (Recombination). This task probes discourse-level coherence in scientific writing. We reuse the 255 three-sentence triplets from Task 1, shuffle each triplet, and ask the model to restore the original order. See 12 for an example. ###### Task 4: Political Sentence Reordering (Recombination). This task probes discourse coherence in policy briefs. We reuse the 255 three-sentence triplets from Task 2, shuffle each triplet, and ask the model to restore the original order. See 13 for an example. ##### Understanding (Tasks 5-6) Tasks 5-6 evaluate sentence-level comprehension and scientific knowledge. Task 5 classifies sentences from scientific papers into five policy brief-relevant categories. Task 6 evaluates scientific knowledge. ###### Task 5: Sentence Classification (Multiple Choice). This task assesses the classification of policy brief-relevant sentences by rhetorical role or content. The dataset includes 1,200 manually verified samples from paper-brief pairs. We curate in three steps: (i) For each pair, use a templated prompt (see 17.7) to generate 15 labeled examples with GPT-o3; (ii) Political experts manually review labels and language quality; (iii) Select a random subset of 1,200 samples. Each instance consists of one sentence and a label from a fixed set (e.g., *Policy Problem*, *Scientific Research Findings*, *Scientific Research Study Methods*, *Policy Implications*, *None*). See 14 for an example. ###### Task 6: Scientific Knowledge Understanding (Multiple Choice). This task evaluates broad scientific knowledge with multiple-choice questions from MMLU-Pro (Wang et al. 2024). The full subset contains 3,511 questions across health, chemistry, economics, and biology. We treat these as one task and sample 1,000 questions randomly for evaluation. This baseline enables comparison between general-purpose and science-specialized models. See 15 for an example. ##### Summarization (Tasks 7-10) Tasks 7-10 evaluate policy-oriented summarization. We show the details as follows. ###### Task 7: Policy Problem Summarization (Writing). We evaluate the ability to identify and summarize the policy problem motivating a study. The dataset includes 200 examples. We curate the dataset following three steps: (i) For each pair, select up to three paragraphs relevant to the brief’s policy problem; (ii) If fewer exist, include all reliably matched paragraphs; (iii) From the final 233 paragraph sets, sample 200 randomly for evaluation. See 16 for an example. ###### Task 8: Research Findings Summarization (Writing). We evaluate the summarization of core scientific findings. The dataset includes 200 examples. We curate the dataset following two steps: (i) For each pair, select three paragraphs aligned with the brief’s findings; (ii) From 255 sets, sample 200 randomly for evaluation. See 17 for an example. ###### Task 9: Study Methods Summarization (Writing). We evaluate the summarization of methodology in policy brief-relevant terms. The dataset includes 200 examples. We curate it following two steps: (i) For each pair, select three paragraphs that describe the methodology reflected in the brief; (ii) From 255 sets, sample 200 randomly for evaluation. See 18 for an example. ###### Task 10: Policy Implications Summarization (Writing). We evaluate the articulation of policy implications grounded in the source paper. The dataset includes 200 examples. We curate the dataset following three steps: (i) For each pair, select up to three paragraphs that support the brief’s implications; (ii) If fewer exist, include all reliably matched paragraphs; (iii) From 222 sets, sample 200 randomly for evaluation. See 19 for an example. ##### Generation (Tasks 11-15) These tasks evaluate brief generation: policy problem, findings, methods, implications, and full brief. We provide a detailed justification for separating section-by-section generation (Tasks 11-14) from full-brief generation (Task 15) in 13.7. In short, Tasks 11-14 complement Task 15 by disentangling factual precision from holistic coherence. Section-level generation emphasizes accurate grounding, while full-brief generation assesses overall readability. Evaluating both provides a more comprehensive view of LLM performance and uncovers trade-offs that Task 15 alone can not capture. ###### Task 11: Policy Problem Generation (Writing). We generate the *Policy Problem* section from the full scientific paper (85 examples). For each pair, we extract the brief’s *Policy Problem* as a reference and provide the full paper as input. For 20 *Journal of Health and Social Behavior* pairs, they lack a clear *Policy Problem* section. We prompt GPT-o3 (see 17.8) to construct the *Policy Problem* section with inputs: the full paper, the brief’s *Research Problem & Data* section, and three in-context expert-written examples from the remaining 65 *Nature* journals. See 20 for an example. ###### Task 12: Research Findings Generation (Writing). We generate the *Research Findings* section from the full paper (85 examples). For each pair, we extract the brief’s *Research Findings* as reference and provide the full paper as input. See 21 for an example. ###### Task 13: Study Methods Generation (Writing). We generate the *Research Study* section (85 examples). For each pair, we extract the brief’s *Research Study* as reference and provide the full paper as input. For 20 *Journal of Health and Social Behavior* pairs, the briefs lack a clear *Research Study* section. We prompt GPT-o3 (see 17.9) to construct the *Research Study* section using the full paper, the brief’s *Research Problem & Data* section, and three in-context expert-written examples from the remaining 65 *Nature* journals. See 22 for an example. ###### Task 14: Policy Implications Generation (Writing). We generate the *Policy Implications* section (85 examples). For each pair, we extract the ground-truth policy implications as a reference and provide the full scientific paper as input. See 23 for an example. ###### Task 15: Policy Brief Generation (Writing). We generate an entire policy brief (85 examples). For each pair, we build the reference by concatenating *Title*, *Policy Problem*, *Research Findings*, *Research Study*, and *Policy Implications*. Input is the full paper. See 24 for an example. ##### Verification (Tasks 16-18) Tasks 16-18 assess consistency between claims (findings or policy implications) and the source paper. ###### Task 16: Scientific Claims Verification (Multiple Choice). We evaluate whether a scientific claim is supported by the paper (850 samples). We construct in two steps: (i) For each paper, prompt GPT-o3 (see 17.10) to generate 10 labeled samples (total 850); (ii) Our political experts manually review all samples and correct three issues. See 25 for an example. ###### Task 17: Scientific Claims Verification 2 (Multiple Choice). We evaluate claim-evidence entailment using SciRIFF subsets (Wadden et al. 2024): *covidfact entailment*, *healthver entailment*, and *scifact entailment* (1,220 samples). We sample 1,000 randomly for evaluation. Each instance presents a claim (e.g., *support*) and associated evidence. See 26 for an example. ###### Task 18: Policy Implications Verification (Multiple Choice). We evaluate whether a policy implication follows from the paper (700 samples). We construct in four steps: (i) Extract all implications from each brief and label them *support*; (ii) Prompt GPT-o3 to generate contradicted implications (see 17.11) and label them *contradict* after manual review of our experts; (iii) Combine to yield 706 samples; (iv) Sample 700 randomly for evaluation. Each instance includes a paper, a policy implication, and a label (*support* or *contradict*). See 27 for an example. #### Evaluation Metrics In this part, we report detailed evaluation metrics for each task, with a particular focus on Tasks 11-15. ###### Micro F1 (Tasks 1-6, 16-18). We compute per-item correctness and use Micro-F1 as the main score. We choose this metric for classification tasks with a firm correct answer. This group includes autocompletion (Tasks 1-4), understanding (Tasks 5-6), and verification (Tasks 16-18). ###### Reference-free Score (Tasks 7-10). We use Gemini-2.5-Pro as an LLM judge to score section summaries. The judge evaluates four dimensions of a summary: *clarity*, *accuracy*, *coverage*, and *overall quality*. We select this metric because these tasks involve free-form writing without a single referenced correct answer. See 17.1 for the full prompt and calculation details. For Tasks 11–15, we first demonstrate the limitations of BERTScore and ROUGE scores, and then introduce a task-specific reference-based score for more accurate evaluation. ###### Limitations in BERTScore and ROUGE Scores for Tasks 11-15. As shown in 13.1, BERTScore remains high even when key sections are missing, as overlapping words inflate similarity. ROUGE scores penalize paraphrasing and drop sharply with minor grammatical changes, despite preserved meaning. Neither metric captures reasoning, structure, or evidence linkage. ###### Reference-based Score (Tasks 11-15). We evaluate the generation tasks using content-aware LLM judging, guided by paper-grounded rubrics that rely on both the paper and policy brief sections. ###### Task 11 (Policy Problem). We score by content and structure because a policy problem contains linked and causal sentences. We describe five parts as the full space of content in a policy problem, but any subset may appear, and the order may vary. (i) *Background* sets the scene. (ii) *Existing problem* states the current obstacle. (iii) *Consequences* describe risks if the problem stays unsolved. (iv) *Attention problem* names the issue that calls for action. (v) *Supporting detail* adds facts, numbers, or sources that help this flow. For each part, we judge two things: its *importance* in the paper and its *quality* in the candidate. This checks what to say and how well it is said, balancing relevance and quality. See 17.2 for the full prompt and calculation details. ###### Task 12 (Research Findings). We score by content only because findings are mostly independent. The judge rates five aspects. We check (i) *completeness*, (ii) *importance*, and (iii) *accuracy* of the candidate findings. (iv) *Summarizing findings* checks if the text highlights the key results rather than a long list. (v) *Specification to findings* checks scope, context, and limits. This rubric rewards correct, essential, and well-focused content. See 17.3 for the full prompt and calculation details. ###### Task 13 (Study Methods). We score by content only because method points are independent. The LLM judge rates three aspects. (i) *Clarity and purpose* checks if the text explains what method is used and why, in a clear output. (ii) *Technicality appropriateness* checks if the level of detail fits a policy audience without jargon. (iii) *Explanation of terms* checks if models, data, and acronyms are explained in plain words. Note that *clarity and purpose* and *technicality appropriateness* carry more weight in evaluation, because *explanation of terms* only serves as an extra signal. This rubric rewards clear intent, appropriate detail, and good definitions. See 17.4 for details. ###### Task 14 (Policy Implications). We score by content only because implications are written as separate points. The LLM judge rates four aspects. (i) *Accuracy* checks if the implications are supported by the paper without speculation or hallucination. (ii) *Coverage* checks if all major implications are included. (iii) *Conciseness and distinctness* checks if each implication is concise and non-redundant. (iv) *Alignment with paper intent* checks if the implications match the paper’s main message, such as a recommendation, warning, or call to awareness. This rubric rewards grounded, complete, and actionable implications. See 17.5 for the full prompt and calculation details. ###### Task 15 (Full Policy Brief). We score by content and style together. The LLM judge rates four aspects. (i) *Contextual depth* checks if the brief captures key findings, methods, and context without missing facts or adding fluff. (ii) *Hallucination risk* checks if every claim is traceable to the paper, with penalties for unsupported numbers or causal links. (iii) *Readability tone* checks if the text is concise, structured, active, and suitable for policymakers. (iv) *Actionability* checks if the implications are concrete, tied to evidence, and useful for policy. This rubric rewards briefs that are accurate, clear, credible, and practical. See 17.6 for the full prompt and calculation details. ### Sci2Pol-Corpus In this section, we give the details of the Sci2Pol-Corpus. It comprises 639 high-quality paper-brief pairs curated from 5.6 million policy documents. It includes three steps: (i) Retrieving candidate science-policy pairs (3.1); (ii) Filtering high-quality pairs with an LLM-as-a-judge approach (3.2); (iii) Polishing briefs through in-context revision (3.3).

Overview of the Sci2Pol-Corpus Curation Process. It consists of 639 paper–brief pairs. Pair retrieval relies on the original policy documents and the scientific papers they cite as candidates.
#### Candidate Pair Retrieval We begin with a large-scale political dataset collection. This collection is derived from documents indexed by Overton, the world’s largest database of policy literature(Furnas et al. 2025; Szomszor and Adie 2022; Yin et al. 2022). From there, we retrive pdf and html policy documents from the public websites of the original publishing organizations and the United States Government Printing Office, IGOs like, the OECD, the UN, and the WHO and numerous think tanks like the Brookings Institution and the Heritage Foundation. These documents cover a broad range of domains, topics and geographic regions. They provide a rich foundation for identifying scientific publications cited in real-world policy contexts. Leveraging Overton’s citation metadata, we identify the scientific papers cited by each policy document. Each citation forms a candidate paper-brief pair, and links a scientific publication to a policy document that references it. To prioritize relevance, we apply a heuristic: The fewer scientific papers a policy document cites, the more likely it is to focus on each one. This assumption increases the likelihood that the policy document reflects or interprets the cited scientific content. Based on this insight, we retain only policy documents that cite no more than three scientific papers. This yields a high-quality pool of 140,000 candidate pairs for further filtering. #### LLM-based Quality Filtering In this stage, we employ GPT-o3 for automated quality filtering to assess whether the policy document centers on the scientific content. It extracts 639 high-quality pairs from 140,000 candidate pairs with two steps: (i) Coarse-grained filtering based on the alignment between the scientific paper abstract and the policy document, and (ii) Fine-grained filtering based on the alignment between the full scientific paper and the policy document. The rationale for this design is to reduce the filtering cost. ###### Coarse-grained Filtering. Scientific papers are often long, typically over 10 pages, and sometimes more than 30. Assuming 500 words per page, this corresponds to about 5,000 to 15,000 words per paper. Assuming each word maps to one token in the GPT-o3 embedding layer, the total token count for 140,000 papers exceeds 700 million. At a rate of \$2 per million tokens, the cost of processing full texts becomes very high. To reduce cost while preserving essential information, we use only the paper abstracts in this filtering stage. We extract these abstracts from SciSciNet (Lin et al. 2023). If a paper is not found in SciSciNet, we discard the corresponding candidate pair. Given a scientific paper abstract and its associated policy document, we prompt GPT-o3 to assess whether the policy document centers on the scientific content described in the abstract. See 40 for the detailed prompt of GPT-o3. As a result, we obtain 1,407 potential high-quality pairs from this step. After the GPT-o3-based filtering, we observe that some policy documents are too long and exceed 10 pages. For standard policy briefs, we prefer shorter documents, typically fewer than 10 pages. Among the 1,407 pairs, 777 pairs contain policy documents under 10 pages, while 630 pairs involve longer documents. To make use of the 630 longer policy documents, we manually extract their executive summaries when available and treat these as the pseudo-policy briefs. For each such case, we constructed two pairs: (i) the executive summary paired with the corresponding scientific paper, and (ii) the executive summary paired with the remaining portion of the policy document. In the second case, the remaining text serves as a pseudo-scientific paper, since it often delivers science-related technical details. If a long policy document did not contain a summary, we discarded it. After this process, we retain 234 usable pairs from the long policy documents. Combined with the 777 pairs involving policy documents under 10 pages, this yields a total of 1,011 curated pairs.

Visualization of Performance of 13 LLMs on Sci2Pol Bench. We show the average performance of commercial and open source LLMs across the five categories of the Sci2Pol-Taxonomy.
###### Fine-grained Filtering. We use GPT-o3 with the full scientific paper and the policy document for fine-grained filtering. The main goal is still to verify whether the policy document centers on the scientific content. However, some pairs originate from the same long policy document, where the executive summary serves as a policy brief and the remaining content acts as a proxy for the scientific paper. We need to avoid these pairs if the two texts are too similar. To handle this, we add a new criterion that measures similarity between the paper and the policy document. This step goes beyond the original metrics used in coarse-grained filtering. See 41 for the detailed GPT-o3 prompt. #### In-context Polishing The last step polishes the policy document. Our policy documents come from official organizations and focus on policy, not briefs. Their format and style do not match a standard policy brief. We propose in-context polishing to align them. We select three expert-written paper-brief pairs from the 85 pairs in Sci2Pol-Bench as references. We then provide the scientific paper and the policy document to GPT-o3 and ask it to revise the document into a standard policy brief while preserving facts and citations. See 42 for the detailed prompt. This step does not inject scientific or political content from the three expert-written paper-brief pairs. It only follows their writing style and format. To further validate that this step does not introduce information leakage from Sci2Pol-Bench, we conduct additional experiments, with details provided in 13.9. ### Experimental Studies In this section, we present the performance of LLMs on Sci2Pol-Bench (4.1) and demonstrate the performance gains achieved through supervised fine-tuning on Sci2Pol-Corpus (4.2). #### LLMs Performance on Sci2Pol-Bench We evaluate 13 models, including both commercial and open-source models: (1) ChatGPT-4o, (2) Claude-3.7-Sonnet, (3) Gemma-3-27B, (4) Gemma-3-12b, (5) Grok-3-beta, (6) DeepSeek-R1, (7) DeepSeek-V3, (8) Qwen3-235B-A22B, (9) Qwen3-8B, (10) Mistral-Large (11) LLaMA-4-Maverick-17B-128E, (12) LLaMA-3.3-70B-Instruct, and (13) LLaMA-3.1-8B-Instruct. For each setting, we conduct 1,000-iteration bootstrap significance tests (seed = 42) and report the mean and standard deviation. We summarize average performance scores across the Sci2Pol-Taxonomy in [tab:zero_avg] and visualize the performance of open-source and commercial models in 6. 12.2 provides full results for all 18 tasks, and see 12.1 for the detailed experimental settings. We include further analysis in 13, covering: (1) human baseline (13.2); (2) common failure modes (13.3); (3) what SFT learns (13.4); (4) the impact of prompt length (13.5); (5) the reliability of the Gemini-based judge (13.6); (6) over-endorse analysis in Tasks 16 and 18 (13.8); (7) information leakage testing of in-context polishing (13.9); (8) potential circularity in benchmark construction (13.10); (9) saturation analysis (13.11); (10) trade-offs in generating briefs from abstracts, introductions, or full papers (13.12). #### Supervised Fine-tuning on Sci2Pol-Corpus We validate the effectiveness of Sci2Pol-Corpus by fine-tuning three models: (1) LLaMA-3.1-8B-Instruct, (2) Gemma-3-12B, and (3) Gemma-3-27B. [tab:sft_avg] shows the results. All fine-tuned models show consistent improvements across Sci2Pol-Bench. Notably, fine-tuned Gemma-3-27B outperforms both GPT-4o and DeepSeek-V3 (671B), despite their significantly larger scale or broader capabilities. These findings highlight the value of domain-specific supervision in Sci2Pol-Corpus. It indeed captures policy-relevant reasoning. See 12.1 for the experimental details. The cross-domain generality of our SFT methodology is further shown in 14. ### Conclusion We present Sci2Pol-Bench and Sci2Pol-Corpus, the first benchmark and training dataset designed specifically for evaluating and improving large language models on scientific-to-policy brief generation. Grounded in a five-stage taxonomy that mirrors the human writing process (Autocompletion, Understanding, Summarization, Generation, and Verification), Sci2Pol-Bench offers a fine-grained evaluation framework spanning 18 tasks. Our results show that even state-of-the-art LLMs struggle with factual grounding, actionability, and policy-appropriate reasoning, and that commonly used metrics such as ROUGE and BERTScore fail to capture brief-writing quality. To address these limitations, we curate Sci2Pol-Corpus from 5.6 million policy documents, producing 639 high-quality paper–brief pairs through citation-based retrieval, LLM-assisted filtering, and expert-guided in-context polishing. Fine-tuning on this corpus yields consistent gains across Sci2Pol-Bench. Notably, a fine-tuned Gemma-27B model surpasses much larger commercial systems, demonstrating that targeted, domain-specific supervision can outweigh scale for scientific-to-policy translation. Together, these resources establish foundational infrastructure for studying and improving how language models translate scientific evidence into actionable policy guidance. We hope Sci2Pol-Bench and Sci2Pol-Corpus will catalyze future research toward models that not only understand science but can responsibly and effectively inform policy in high-stakes societal domains. We also provide detailed LLM usage in 6, dataset links in 7, limitations, future work, and broader impacts in 9, and related works in 10. ### Reproducibility Statement We have made efforts to ensure the reproducibility of our work. The full details of our benchmark construction, dataset curation, and model evaluation are provided in the main text and appendices. For all 18 tasks in Sci2Pol-Bench, we describe the data sources, annotation protocols, and evaluation metrics in detail (2). The construction steps of Sci2Pol-Corpus, including citation-based retrieval, LLM-based filtering, and in-context polishing, are documented in 3. We release all prompt templates and scoring rubrics in 17. To facilitate replication, we include the code in the supplementary materials. For each supervised fine-tuning experiment, we specify training configurations and hardware setup in 12.1. Collectively, these resources ensure that our results can be verified and extended by the community. ### Ethic Statement ###### Human-in-the-loop Annotation Process. All reviewers involved in task development and verification are postdoctoral researchers or research faculty in political science. They are co-authors of this paper and serve as domain experts throughout the benchmark and dataset construction. No monetary compensation is provided; their participation is motivated by scholarly collaboration and a shared commitment to advancing evidence-based policymaking. ###### Details of Human Involvement in Label Generation.  Three tasks (5, 16, and 18) use GPT-o3 to propose initial labeled items, and all of them undergo full human verification. Each task is reviewed independently by two reviewers, with 100% reviewer agreement across all accepted items. For all 1,200 samples in Task 5, no corrections are required. For Task 16, reviewers examine all 850 samples and correct three mislabeled cases (correction rate $`\approx 0.35\%`$). For Task 18, reviewers validate all 706 generated contradictory implications, and no corrections are needed. ###### Copyright and Use of Scientific Papers. The benchmark and dataset are built from publicly accessible content such as article metadata. We release only article DOIs, metadata, access dates, policy brief metadata, and our annotations. The benchmark and dataset are released under CC-BY-NC 4.0, with a two-tier protocol (open metadata + controlled transformed excerpts), and no publisher-owned text is redistributed. Derived tasks (e.g., summarization in Sci2Pol-Bench) rely solely on reformulated excerpts for non-commercial academic purposes. We also document potential impacts and implement explicit safeguards: all tasks require evidence-based summarization (not prescriptive advice), prompts prohibit unsupported claims, and evaluation judges penalize fabricated, speculative, or over-confident policy recommendations. To reduce evaluation variance and bias, we employ rubric-anchored scoring, multi-model judging, and expert-labeled calibration samples. These measures establish clear licensing boundaries, release protocols, bias-mitigation procedures, and risk-assessment practices appropriate for policy-oriented datasets. For transparency, we provide a detailed list of all used paper–brief pairs in 15.1, including DOIs, sources, and access dates. ### Acknowledgments WW would like to thank Robin Luo, Yichi Zhang, Qineng Wang, and Kongdi Zheng for enlightening discussions on related topics. We acknowledge grant support from the National Science Foundation (award No. **2404035**). # Supplementary Material ### LLM Usage Acknowledgement Large Language Models (LLMs) were used in this project as general-purpose assistive tools across several stages of data curation, benchmark design, evaluation, and paper writing. We detail their precise contributions below. - **Manuscript Editing:** We used LLMs (e.g., **GPT-4o**) to improve the clarity, grammar, and fluency of English writing throughout the paper. - **Sci2Pol-Bench:** (i) Data Augmentation and Annotation: For Understanding, Generation, and Verification tasks, **GPT-o3** was used to generate training and evaluation samples, including sentence classification data, missing brief sections (e.g., Policy Problem and Study Methods for JHSB samples), and supported/contradicted claims for claim verification tasks. (ii) Prompt-Based Evaluation: **Gemini-2.5-Pro** was used as an LLM-as-a-judge to evaluate model outputs for Tasks 7–15 based on clarity, factual accuracy, completeness, and policy relevance. (iii) LLM Evaluation: We benchmarked 13 commercial and open-source LLMs. - **Sci2Pol-Corpus:** (i) Filtering and Evaluation of Candidate Pairs: We employed **GPT-o3** to identify 639 high-quality science-policy pairs from 140,000 candidates. (ii) In-Context Polishing of Policy Briefs: We used **GPT-o3** with three expert-written examples in an in-context setup to revise raw policy documents into standard, structured policy briefs. All LLM-generated content used in data or evaluation workflows was curated or reviewed by human experts. We take full responsibility for all content in this paper, including that generated or evaluated by LLMs. No LLM is credited as an author. ### Dataset Links We provide the links for the datasets: - Sci2Pol-Bench Dataset: \ - Sci2Pol-Corpus Dataset: ### Preliminaries In this section, we present preliminaries on the policy brief (8.1) and compare it with the full scientific paper, paper introduction, and paper abstract (8.2). #### Policy Brief The policy brief is first introduced by *Nature Energy* to provide concise, policy-relevant summaries written by the original researchers. Each brief cites its source article using clear metadata (e.g., “based on: title doi”), enabling reliable paper-brief alignment. It includes five parts: - **Title:** A concise one-line headline that states the policy issue. - **Policy Problem:** A short paragraph (fewer than five sentences) framing the societal risk. - **Scientific Research Findings:** One or two compact paragraphs (about 150 words) clearly summarizing the study’s core quantitative results and key empirical insights. - **Scientific Research Study Methods:** A single paragraph (about 100 words) briefly explaining the dataset and modelling approach in clear, accessible, and lay terms. - **Policy Implications:** Four to six bullet points (about 25 words each), drawn strictly from the paper’s authors, highlighting concrete conclusions directly relevant to policymakers. No added recommendations, speculation, or external examples. #### Comparison with Full Scientific Paper, Paper Introduction and Abstract We provide a comparison between the policy brief and the full paper, introduction, and abstract. - **Purpose:** A research paper provides full evidence and methods for experts; its introduction frames the knowledge gap and aims; the abstract compresses aims, methods, and key results. A policy brief removes technical detail and distills only policy-relevant insights for decision makers. - **Structure:** Papers follow IMRaD (Introduction, Methods, Results, Discussion) with references; abstracts compress these into one paragraph. Policy briefs replace IMRaD with Policy Problem, Scientific Research Findings, Scientific Research Study Methods, and Policy Implications. - **Tone and Jargon:** Papers, introductions, and abstracts use technical language and equations. A policy brief avoids jargon and equations, relying on plain prose accessible to non-experts. - **Focus:** Full papers emphasize evidence and methodological rigor; introductions stress scholarly significance; abstracts distill what is done and found. Policy briefs highlight why findings matter for policy, and repeat only the author’s stated implications. - **Audience Takeaway:** Researchers consult full papers for replicable detail, introductions for rationale, and abstracts for a quick overview. Policymakers rely on briefs to grasp the problem, key evidence, methodological credibility, and policy implications. ### Limitations, Future Work, and Broader Impact We present the limitations and future work (9.1), and the broader impact (9.2). #### Limitations and Future Work We have the following limitations and related future work: - Sci2Pol-Bench and Sci2Pol-Corpus remain modest in size. Unlike efforts that rely on synthetic summaries or crowd-sourced approximations, our datasets reflect the communicative intent and expertise of scientists writing for policy audiences. We will continue to incorporate newly published briefs and their corresponding papers to expand coverage. - Although supervised fine-tuning on Sci2Pol-Corpus yields consistent gains, it does not yet surpass state-of-the-art commercial models, e.g., Grok. Future work should explore improved training strategies and better use of Sci2Pol-Corpus to close this gap. - The existing 85 published policy briefs come from *Nature Energy*, *Nature Climate Change*, *Nature Cities*, *Nature Sustainability*, and the *Journal of Health and Social Behavior (JHSB)*. Expanding Sci2Pol-Bench and Sci2Pol-Corpus to additional disciplines and languages will enable broader generalization as new paper-brief pairs appear. - 20 samples in Tasks 11 and 13 are drawn from the *Journal of Health and Social Behavior (JHSB)*. For these cases, the *Policy Problem* and *Study Methods* sections are revised by GPT-o3 based on the *Research Problem & Data* section. This step may not be as reliable as for the remaining 65 samples. However, our experts review them, and we chose to retain these cases because they account for more than 20% of the total sample set. - While Sci2Pol-Bench is intentionally designed around single-paper policy briefs, this focus also imposes an inherent limitation. Our benchmark does not capture the broader genre of multi-source policy synthesis, where policymakers integrate evidence from many studies. As a result, our benchmark does not assess, nor does our dataset enhance, a model’s capacity to aggregate heterogeneous sources or integrate multiple lines of evidence. #### Broader Impact This work advances understanding of the critical connection between scientific research and policymaking. Science provides evidence and authoritative knowledge essential for informed decisions and sustaining public trust. Its role grows as pressing challenges such as climate change, public health crises, and rapid technological change demand the timely integration of new findings. Sci2Pol-Bench directly addresses a persistent bottleneck in this pipeline: policymakers’ difficulty in translating dense, nuanced scientific evidence into actionable guidance. By decomposing the brief writing process into five stages and providing 18 fine-grained tasks, it offers the first systematic evaluation framework tailored to scientific-to-policy communication. This benchmark exposes weaknesses in even frontier language models and creates opportunities to build models that are not only accurate but also clear, consistent, and persuasive for policy audiences. Sci2Pol-Corpus complements this evaluation framework by supplying the first domain-specific training dataset for policy brief generation. Curated from millions of policy documents and refined through quality filtering and in-context revision, it provides 639 high-quality papers to brief pairs. Fine-tuning on this corpus leads to consistent gains across Sci2Pol-Bench, even enabling smaller open source models to surpass frontier-scale models. This demonstrates the importance of targeted supervision for capturing the communicative intent of scientific-to-policy translation. Together, these resources have a significant and lasting societal impact worldwide. They empower scientists, policymakers, and institutional leaders with practical tools to measure and improve the policy relevance of scientific communication. They strengthen accountability and reduce serious risks of misinformation by encouraging accurate, structured, and transparent communication of evidence. They also highlight the vital and growing role of science in shaping societal outcomes and help ensure that policy decisions remain grounded in rigorous research rather than speculation. At a time of political polarization, resource constraints, and global competition, the ability to generate clear, accurate, and actionable policy briefs is vital. Sci2Pol-Bench and Sci2Pol-Corpus provide foundational infrastructure for developing models that meet this challenge, supporting evidence-based decision-making and enhancing public trust in science and governance. ### Related Works We review related work on benchmarks and datasets in the scientific and political domains. #### Scientific and Political Benchmarks In recent years, LLMs (Team et al. 2024; A. Liu et al. 2024; **touvron2023LLaMA?**; Achiam et al. 2023; Bai et al. 2023) have attracted significant attention due to their impressive performance. There have been several benchmarks to evaluate their performance in scientific and political domains. ###### Scientific Benchmarks. In the scientific domain, benchmarks such as SciRIFF (Wadden et al. 2024), MMLU-Pro (Wang et al. 2024), SciInstruct (D. Zhang et al. 2024), SciLitLLM (S. Li et al. 2024), and SciRepEval (Singh et al. 2023) test models on tasks like summarization, question answering, and claim verification. These benchmarks focus on instruction-following and comprehension of scientific content. Most use single-step tasks with scientific inputs, typically full papers or extended passages. ###### Political Benchmarks. In the political domain, benchmarks assess how models reason about political science, ideology, and value alignment. Political Science QA (L. Li et al. 2024) tests factual knowledge and reasoning. Röttger et al. (2024) and Ren et al. (2024) probe value orientation and political opinions. Motoki et al. (2024) measure political bias. These tasks often use short prompts under 300 tokens and rely on multiple-choice formats. Political-LLM (L. Li et al. 2024) contributes a useful taxonomy and discussion of use cases, but offers no benchmark or annotated data. However, existing scientific and political benchmarks overlook the dual challenge of science-informed policy communication: understanding complex research and translating it into actionable language. SciRIFF and SciInstruct focus on instruction-following for scientific tasks, while MMLU-Pro tests expert-level reasoning with multiple-choice questions. Political benchmarks such as ValueBench or Political-LLM probe ideology and values but do not address policy generation or scientific grounding. To fill this gap, Sci2Pol-Bench evaluates LLMs on generating policy briefs from full-length scientific papers, pairing them with expert-written briefs and introducing a five-stage pipeline. We compare against three representative efforts. SciRIFF and MMLU-Pro are the closest scientific benchmarks: SciRIFF covers single-step tasks like summarization, QA, and claim verification, while MMLU-Pro measures reasoning breadth but not policy translation. Political-LLM provides the first principled framework for computational political science, offering a taxonomy but no annotated tasks. Together, these works form the strongest prior art in scientific comprehension and political reasoning. Sci2Pol-Bench builds on them but uniquely combines scientific fidelity with direct policy relevance, making it the first benchmark to evaluate LLMs on producing accurate, actionable policy briefs. #### Scientific and Political Datasets Beyond evaluation benchmarks, several open-source fine-tuning datasets adapt LLMs to scientific and political domains through supervised instruction tuning or continued pre-training. ###### Scientific Datasets. Researchers have curated domain-specific corpora to enhance scientific reasoning. SciInstruct (D. Zhang et al. 2024) provides curated instructions across physics, chemistry, math, and formal proofs, improving models’ performance on college-level problems. SciLitLLM (S. Li et al. 2024) combines continual pre-training on research papers with supervised tuning. It introduces SciLitIns, a collection of instructions targeting underrepresented fields. Large-scale corpora have also been used. For example, Galactica (Taylor et al. 2022) is trained on 48 million science documents. This shows how domain-specific training endows models with broad scientific knowledge. Together, these datasets demonstrate the effectiveness of specialized fine-tuning. ###### Political Datasets. In the political domain, fine-tuning datasets emphasizes factual knowledge, ideology, explicit opinions, and bias alignment. PoliTune (Agiza et al. 2024) creates ideology-specific instruction data from social media, while Vendetti et al. (2025) fine-tune the existing LLMs on curated Reddit discourse to model polarized political commentary. These efforts show how targeted fine-tuning can steer models toward particular political knowledge or stances. However, existing datasets in both domains overlook the dual challenge of science-informed policy communication. Scientific datasets like SciInstruct and SciLitIns strengthen technical reasoning but exclude crucial policy text, while political datasets such as PoliTune emphasize ideological alignment yet ignore underlying scientific evidence. Sci2Pol-Corpus fills this important gap by linking full-length research papers to expert-written policy briefs. It trains advanced models to extract key findings and translate them into clear, actionable recommendations for diverse policymakers. ### How LLMs Fail in Policy Brief Generation In this section, we present representative examples that highlight the limitations of existing LLMs in scientific-to-policy brief generation. Using the prompt in 24, we instruct different LLMs to generate policy briefs from the scientific papers listed in 15, and then compare their outputs with the expert-written published briefs. The results reveal issues in four key areas: contextual depth, hallucination risk, readability and tone, and actionability. We provide one illustrative example for each issue in 11.1, 11.2, 11.3, and 11.4. #### Contextual Depth 1 is an example where LLM output contains a contextual-depth issue. The issue is evident because the published brief gives precise quantitative findings such as “an additional 0.5 GtCO2/yr by 2030 and 1.9 GtCO2/yr by 2050” and compares these against Paris Agreement scenarios to highlight the “CDR gap”. However, the LLM output only provides a broad summary about reliance on land-based removals and a lack of novel technologies. It does not capture the specific numerical evidence and scenario-based reasoning that are central to the paper’s findings.
| **Expert-written Brief** | **LLM Output** | |:---|:---| | In our study we found that compared with 2020, the most ambitious national proposals for CDR imply an additional 0.5 GtCO2yr-1 of removals by 2030, and 1.9 GtCO2yr-1 by 2050. Compared with CDR scaling in Paris Agreement-consistent scenarios, we found that these national CDR proposals tend to fall short by hundreds of megatonnes of carbon dioxide in 2030 to several gigatonnes of carbon dioxide in 2050, highlighting a ‘CDR gap’. However, we find that the most ambitious proposals do come close to levels in a low-energy-demand scenario where CDR requirements are minimized, suggesting that if countries pledge more ambitious emissions reductions consistent with these scenarios, the CDR gap will be closed. As levels of reporting vary, our evaluation of proposed CDR does assume that a number of countries simply maintain their current levels of (conventional) removals. In addition, it remains unknown to what extent firm CDR policies will follow these proposals. | Research indicates that current national proposals for CDR are not aligned with integrated assessment model scenarios, which project the levels necessary to achieve the 1.5°C temperature limit. The study shows that while afforestation and land-based removal methods are in use, there is a lack of commitment to scaling novel CDR technologies like bioenergy carbon capture and storage (BECCS) or direct air carbon capture and storage (DACCS). Most proposed CDR efforts remain heavily reliant on land-based removals, which are not sufficient on their own to meet the required targets. | **Expert- vs. LLM-written Policy Brief Excerpts: Contextual Depth Issue.** Paper: *The carbon dioxide removal gap* (Lamb et al. 2024b). LLM: GPT-4o. Presented Brief Section: Scientific Research Findings.
#### Hallucination Risk 2 shows a risk of hallucination. In this case, the LLM reads the paper as endorsing public or cooperative ownership, progressive tax swaps, cross-border subsidies, and profit-cap rules. The paper does not state these. The human brief stays close to the source and lists the core points: trillion-dollar scale, windfall profits under unregulated markets, trade-offs from profit caps, regional distribution effects, and the near-term need to scale NETs. The LLM adds new policies and causal claims, so it may mislead readers.
Expert- vs. LLM-written Policy Brief Excerpts: Hallucination Issue. Paper: Inequality repercussions of financing negative emissions (Andreoni et al. 2024b). LLM: Qwen3-8B. Presented Brief Section: Policy Implication.
Expert-written Brief LLM Output

  • In a net-zero emissions world, NETs could become a trillion-dollar business globally.

  • If financed through an unregulated carbon market, the owners of these companies would enjoy windfall profits, potentially leading to a large increase in economic inequality.

  • Market regulation, such as profit caps, could reduce the inequality increase, but at the risk of stimulating too much or too little carbon removal - policymakers should be aware of this trade-off.

  • Concentrating removal efforts in the Global North or transferring resources to the Global South could, to some extent, offset the increase in inequality at the global level.

  • These dynamics mostly apply to a net-zero and post-net-zero world. The current priority of policymakers should remain to provide adequate resources to scale up NETs towards technology maturity.

  1. Equitable Ownership Structures: Policymakers should consider public ownership or cooperative frameworks for NETs to prevent disproportionate benefits to private capital holders. This could mitigate the ownership effect and ensure that profits from carbon removal are distributed more fairly.

  2. Progressive Taxation and Redistribution: Replacing regressive carbon taxes with progressive income taxes after net-zero, or using redistribution mechanisms, could counteract inequality risks. Pre-net-zero taxation reforms might also reduce adverse impacts, though they could compromise economic growth.

  3. International Financial Transfers: Implementing cross-border carbon removal credits or subsidies for low-income countries could address global inequality. However, these transfers may still benefit high-income groups within recipient nations, necessitating targeted redistribution policies.

  4. Regulating Profit Margins: Introducing profit caps or alternative pricing schemes for NETs within a single carbon market could balance incentivizing deployment with equity. Yet, such measures must avoid distorting markets or discouraging investment.

  5. Integrated Policy Design: Climate strategies must explicitly account for distributional impacts in policy frameworks. This includes designing carbon pricing, ownership rules, and revenue allocation mechanisms to align with equity goals while maintaining the viability of NETs as part of a low-carbon transition.
#### Readability and Tone 3 is an example where the LLM output has readability and tone issues. The expert-written brief is logically ordered, uses active voice, and speaks to policymakers. It defines the study design, maps three strategies to plain descriptions, and explains how the team built and compared scenarios. By contrast, the LLM output contains unexplained jargon in bullet points. It also introduces acronyms without setup, which raises the barrier for non-expert readers.
Expert- vs. LLM-written Policy Brief Excerpts: Readability and Tone Issue. Paper: Demand-side strategies enable rapid and deep cuts in buildings and transport emissions to 2050 (Heerden et al. 2025b). LLM: DeepSeek-V3. Presented Brief Section: Scientific Research Study Methods.
Expert-written Brief LLM Output
Our study uses quantitative scenarios generated with global integrated assessment models (IAMs) that are specifically designed to capture the interactions within and between the energy system, the economy, and the environment. We developed three scenarios that each capture the impact of a specific strategy targeting energy use. In the first scenario end users reduce or change energy‑using activities, in the second they adopt more efficient technologies, and in the third they switch to electrification or alternative energy carriers as substitutes to fossil fuels. Using a structured questionnaire, we also gathered inputs from policymakers and experts in demand sectors to refine the scenarios. We quantitatively interpreted and quantified the scenarios through 2050 using seven IAMs. Finally, we compared the resulting emission reductions per strategy across the set of models and scenarios and analysed the results to better understand the main factors driving these reductions.

The study used an ensemble of seven global integrated assessment models (IAMs) to simulate three intervention strategies:

  1. Activity-focused (ACT): Policies promoting shared mobility, reduced travel, and smaller living spaces.

  2. Technology-optimizing (TEC): Enhanced building insulation, vehicle efficiency standards, and retrofits.

  3. Electrification-focused (ELE): Wide-spread adoption of heat pumps, electric vehicles, and renewable energy.

Scenarios were compared against a “current policies” baseline and a 1.5°C-aligned pathway, accounting for regional variations and model uncertainties.
#### Actionability 4 shows weak actionability. The LLM states broad aims: “reconsider approaches”, “build public support”, and “prioritize durable reforms”. It does not say who does what, when, or how to track success. The expert-written brief gives steps and targets: start with the most damaging fuels, especially coal; invest in transit and support EVs; use fuel-efficiency rules; and choose measures that survive political shocks. The LLM repeats goals but omits scope, sequence, and levers, so a policymaker cannot act.
Expert- vs. LLM-written Policy Brief Excerpts: Actionability Issue. Paper: Fossil fuel subsidy reforms have become more fragile (Mahdavi et al. 2025a). LLM: Gemma3-12B. Presented Brief Section: Policy Implication.
Expert-written Brief LLM Output

  • While fossil fuel subsidies are harmful, efforts to remove them fail more than 90% of the time. Low-cost fuel is popular and hard to change.

  • Reforms are only beneficial if they endure. While price reforms are fragile, other kinds of policies — such as investments in clean transportation and support for electric vehicles — are more durable.

  • Since subsidies are politically difficult to remove, governments should concentrate their efforts on the most damaging types of fuel, especially coal.

  • Governments should look for alternative ways to reduce the demand for subsidized fuel, for example, by using regulatory measures to enhance fuel efficiency, investing in public transit and promoting electric vehicles.

The findings suggest that current strategies for reducing fossil fuel subsidies are failing. Policymakers should reconsider their approaches, prioritizing:

  • Long-Term Strategies: Focus on policies that reduce demand for subsidized fuel (e.g., fuel efficiency standards, electric vehicle incentives, improved public transit) alongside, rather than in place of, subsidy reductions.

  • Political Feasibility: Recognize the political challenges of subsidy reform and build public support through clear communication, targeted compensation for vulnerable populations, and a gradual, phased approach.

  • Durable Reforms: Prioritize reforms like shifting to floating prices that are not easily eroded by inflationary pressures or exchange rate fluctuations, even if initially less popular.
### Experimental Details We provide: (i) the detailed experimental settings for LLM evaluation on Sci2Pol-Bench and supervised fine-tuning (12.1); (ii) the performance of 13 LLMs across all Sci2Pol-Bench tasks (12.2); and (iii) the supervised fine-tuning results of three LLMs (12.3). #### Detailed Experimental Settings We present the detailed experimental settings. ###### Evaluate LLMs on Sci2Pol-Bench. We evaluate 13 LLMs through API calls. For Grok, GPT-4o, and Claude, we use their respective private APIs. For all other models, we rely on the Novita API. ###### Supervised Fine-tuning. Our supervised fine-tuning uses LLaMA-Factory (Zheng et al. 2024). We apply lightweight Low-rank Adaptation (LoRA) (Shen et al. 2022) under the following setup. The base models are LLaMA-3.1-8B-Instruct, Gemma-3-12B-IT, and Gemma-3-27B-IT. LoRA uses rank 8, alpha 32, and dropout 0.05, applied to the query and value projection matrices of all attention layers. Training runs for six epochs with a batch size of 8, a cosine learning rate schedule, a peak learning rate of 1e-4, and a 5% warmup phase. FlashAttention-2 (Dao 2024) and the Liger kernel (Hsu et al. 2025) are enabled for efficient long-context training. We split Sci2Pol-Corpus into 95% training and 5% validation, and select the checkpoint with the lowest validation loss. We run all experiments on 4 NVIDIA A100 80GB Tensor Core GPUs. #### LLM Performance on Sci2Pol-Bench We show the detailed performance of 13 LLMs across all tasks in [tab:zero_shot_1_recalculated]. Our experiments lead to three main findings: - Current LLMs struggle to generate high-quality policy briefs from scientific papers, even though they likely encounter academic content during pretraining. These results reveal a persistent gap between general language capability and task-specific policy reasoning. - Sci2Pol-Bench exposes new weaknesses not captured by traditional LLM benchmarks. As shown in [tab:zero_avg] and [tab:zero_shot_1_recalculated], tasks in Autocompletion and Understanding produce the widest F1 range across models, highlighting instability in predicting coherent and grounded continuations. - Larger models generally perform better, but size alone doesn’t guarantee robustness. For instance, Gemma-3-12B outperforms the larger Gemma-3-27B on several tasks, e.g., Task 2 (see [tab:zero_avg] and [tab:zero_shot_1_recalculated]). Similarly, commercial models that dominate benchmarks like GPQA (Phan et al. 2025) often underperform on early-stage Sci2Pol tasks such as Task 1 and Task 4. #### Supervised Fine-tuning on Sci2Pol-Corpus We report the detailed supervised fine-tuning performance of three LLMs in [tab:detail_sft]. ### Additional Experimental Analysis In this section, we examine the following aspects in Sci2Pol-Bench and Sci2Pol-Corpus. (1) We provide evidence for the limitations of BERTScore and ROUGE when applied to Tasks 11-15 (13.1). (2) We provide the human baseline for Sci2Pol-Bench (13.2). (3) We provide the analysis of common failure modes of top LLMs (13.3). (4) We provide a detailed analysis of what supervised fine-tuning learns (13.4). (5) We test the impact of prompt length on performance for Tasks 1-4 (13.5). (6) We assess the reliability of the Gemini-2.5-Pro-based reference-free judge by alignwing with human evaluation (13.6). (7) We contrast section-by-section versus full-brief generation to explain the need for Tasks 11-14 in addition to Task 15 (13.7). (8) We analyze whether models show a tendency to over-endorse by studying the distribution of false positives and false negatives in Tasks 16 and 18 (13.8). (9) We validate that in-context polishing does not introduce information leakage (13.9). (10) We evaluate potential circularity in benchmark construction by comparing GPT and DeepSeek families on Task 16 (13.10). (11) We do the saturation analysis (13.11). (12) We compare brief generation from abstracts, introductions, and full papers to analyze trade-offs in context length (13.12). Together, these studies clarify the robustness and fairness of our Sci2Pol-Bench and Sci2Pol-Corpus. #### Limitations of BERTScore and ROUGE Scores for Tasks 11-15 We provide evidence for why BERTScore and ROUGE scores fail to evaluate Tasks 11-15. Consider one example: the LLaMA-3.1-8B-Instruct-generated policy brief for the scientific paper *How Central Banks Address Climate and Transition Risks* (Paper (Shears et al. 2025b); Brief (Shears et al. 2025a)). For BERTScore, we demonstrate that deleting large portions of the candidate brief hardly changes the score against the true brief. Using 5, we define “brief completeness” by removing sections of the candidate brief to create 75%, 50%, 25%, and title-only variants. [tab:bert-rouge-task16] then reports the BERTScore trend as completeness falls. For ROUGE scores, we show that simple grammatical or paraphrasing changes cause the scores to drop sharply, even when meaning is preserved. [tab:policy_paraphrase_texts] presents example texts, and 6 reports ROUGE-1/2/L scores between them. The results reveal low scores despite semantic equivalence, highlighting ROUGE’s sensitivity to surface form. Together, these findings illustrate two failure modes: BERTScore remains high under major deletions, while ROUGE collapses under harmless paraphrases. Consequently, for Tasks 11-15, we rely on task-specific LLM-judge scores, which verify section coverage, reasoning flow, and evidence linkage.
| **Completeness** | **Policy Problem** | **Findings** | **Methods** | **Policy Implications** | |:---|:--:|:--:|:--:|:--:| | **100% (Full)** | | | | | | **75%** | | | | | | **50%** | | | | | | **25%** | | | | | | **Title Only** | | | | | **Deriving Completeness Subsets from LLM Outputs.** Sections are removed sequentially in the order: Policy Implications $`\rightarrow`$ Methods $`\rightarrow`$ Findings $`\rightarrow`$ Policy Problem. The table indicates which sections remain at each completeness level. Percentages reflect the proportion of content retained.
| **Brief Completeness** | **BERT Precision** | **BERT Recall** | **BERT F1 Score** | | | | |:---|:--:|:--:|:--:|:--:|:--:|:--:| | **Full Brief** | 0.8689 | 0.8599 | 0.8644 | | | | | **75%** | 0.8721 | 0.8612 | 0.8666 | | | | | **50%** | 0.8829 | 0.8551 | 0.8688 | | | | | **25%** | 0.8828 | 0.8174 | 0.8489 | | | | | **Title Only** | 0.8738 | 0.7839 | 0.8264 | | | |
Original Text Grammar-alternated (Rephrased) Text

Policy Implications

  • Central banks vary substantially in the extent to which they re-risk stranded asset and physical climate risks and de-risk clean energy investments.

  • Central bank actions on climate risks are positively associated with their country’s climate policy stringency and public concern with climate change and less with its underlying economic risks.

  • Despite their autonomy, central banks do not substitute for the lack of national climate policy but complement existing national policies promoting the clean energy transition.

  • The political nature of central bank actions to manage transition and physical risks raises concerns about unmanaged risks in the global economy, specifically stranded asset risks.

  • A central bank climate index could increase transparency of the risk mitigation gap; international institutions governing central banks could set standards for climate and transition risk management.

Policy Implications

  • Substantial variation exists across central banks in how they re-risk stranded asset and physical climate risks while de-risking clean energy investments.

  • Their actions on climate risks are linked more strongly with national climate policy stringency and public concern over climate change, and less strongly with economic fundamentals.

  • National climate policy is not replaced by central banks, even with their autonomy; rather, it is supported and complemented in advancing the clean energy transition.

  • The political character of central bank efforts to handle transition and physical risks raises concerns about unmanaged threats to the global economy, especially stranded asset risks.

  • Transparency of the risk-mitigation gap could be improved by a central bank climate index, and international institutions governing central banks could set standards for climate and transition risk management.
| **Pair** | **ROUGE-1 F1** | **ROUGE-2 F1** | **ROUGE-L F1** | |:---------------------------|:--------------:|:--------------:|:--------------:| | **Original vs. Rephrased** | 0.4058 | 0.1606 | 0.2319 | **ROUGE on Original vs. Grammar-alternated Paraphrase.** We report ROUGE-1/2/L F1 scores between the two columns in [tab:policy_paraphrase_texts]. Despite identical meaning, scores are low. These illustrate ROUGE’s sensitivity to word order and phrasing rather than semantic equivalence.
#### Human Baseline for Sci2Pol-Bench We recruit two university-affiliated evaluators and compensate each with \$100. Because full evaluation is time-intensive, we do not assess all questions. Instead, we randomly sample 25 questions from each task in Tasks 1–10 and 16–18. For Tasks 11–15, we randomly sample three questions each, using the same three underlying scientific papers across all five tasks to maintain consistency. Due to the limited number of evaluation questions, we report only the final results and do not conduct bootstrap significance testing. The resulting evaluations, together with the state-of-the-art Grok-3-beta results, are shown in [tab:human_testing]. These results indicate that current LLMs still have meaningful room for improvement. #### Analysis of Common Failure Modes In this section, we summarize the most salient failure modes of strong general-purpose LLMs by task group and connect them to our design targets. ###### Autocompletion (Tasks 1–4). On scientific and policy autocompletion and sentence reordering, leading models often rely on shallow lexical cues rather than discourse structure. For Tasks 1–2, models frequently choose continuations that share surface overlap (entities, topical words) but violate local rhetorical flow (e.g., jumping to limitations or policy implications prematurely). For Tasks 3–4, they can correctly identify the “introductory” sentence but struggle to order the remaining sentences when all share similar vocabulary. These failures reflect limited pattern alignment to domain-specific discourse patterns and weak implicit modeling of how scientific and policy arguments unfold. Our design, which exposes models to well-structured paper-to-brief mappings, is designed to strengthen discourse-sensitive pattern alignment and implicit skills for ordering and continuations in this genre. ###### Understanding (Tasks 5–6). For sentence-level classification (Task 5), frontier models often conflate *Policy Implications* with *Recommendations*, or mislabel methodological descriptions as findings when they contain numerical results. This indicates insufficient specialization to Sci2Pol rhetorical roles and weak style awareness of policy-brief section functions. In contrast, performance on general scientific knowledge (Task 6) is comparatively strong and less affected by our domain-specific SFT; we treat Task 6 as a sanity check that specialization to Sci2Pol does not destroy broad knowledge. Our corpus, built around explicit section labels and section-specific prompts, directly targets improved pattern alignment and stylistic control over these roles. ###### Summarization (Tasks 7–10). For policy-oriented summarization, leading models systematically underperform along two axes. First, failures in factual grounding: models often hallucinate policy framings, quantitative effects, or stakeholder groups that are not present in the selected paragraphs, especially in Tasks 7 (Policy Problem) and 10 (Policy Implications). Second, failures in style and pattern alignment: generated outputs drift toward generic “science communication” prose, ignoring the requested section type (e.g., mixing methods into findings, or inserting recommendations in the policy problem). Tasks 8–9 reveal additional issues with functional-length adaptation: models either compress complex methods and findings into vague topical summaries, or over-elaborate with boilerplate explanations not grounded in the input text. Our in-context polishing prompt and SFT objective explicitly penalize these behaviors by requiring section-specific outputs, professional policy-brief style, and strict factual alignment with the target paper. ###### Generation (Tasks 11–15). When asked to generate sections or full briefs from the full paper, strong models exhibit compounding errors. For Tasks 11–14, they often produce sections that (i) are not cleanly separated, (ii) rely on generic policy tropes rather than paper-specific evidence, or (iii) introduce external context and speculative recommendations beyond the study’s scope. These reflect weaknesses in factual grounding, style discipline, and implicit skill at mapping paper structure into the four-section brief template. For full-brief generation (Task 15), models sometimes maintain local coherence within paragraphs but fail at global pattern alignment: section ordering is irregular, headings are omitted or merged, and later sections contradict earlier claims. Our design addresses these issues by (i) training on high-quality, section-aligned paper–brief pairs, (ii) enforcing a stable four-part template in the polishing prompt, and (iii) emphasizing functional length and section-specific content during optimization, thereby improving both pattern alignment and implicit structural skills. ###### Verification (Tasks 16–18). In verification tasks, frontier models tend to over-accept plausible-sounding claims, leading to high recall but poor precision. For Task 16, models often label partially supported or overstated claims as fully supported, revealing limited sensitivity to scope qualifiers and caveats in the source paper. For Tasks 17–18, they struggle to distinguish subtly contradicted policy implications from valid ones, especially when the implication is globally reasonable but locally misaligned with the provided evidence. These failure modes are rooted in weak factual grounding and insufficient disciplined reasoning about what is actually entailed by the text. While Sci2Pol SFT is not trained directly on verification labels, the polishing objective, which enforces strict factual alignment and discourages speculative or extrapolatory language in generated briefs, is designed to mitigate these tendencies by cultivating more conservative, evidence-driven behaviors that transfer to verification settings. #### Analysis of Improvements from SFT In this section, we analyze the improvements induced by supervised fine-tuning (SFT). Based on the design of the in-context polishing prompt (42), we identify four aspects of learning: factual grounding, style, pattern alignment, and implicit skill acquisition. A detailed comparison of base and fine-tuned outputs is provided in Table [tab:qualitative_example_peer_referral], with additional analyses of leading-model shortcomings in [tab:deepseek_limitations,tab:gpt4o_limitations]. **Factual Grounding.** The in-context polishing prompt explicitly requires “strict factual alignment” with the target scientific paper and provides multiple high-quality examples. This setup strongly encourages the model to improve factual grounding. Specifically, the model learns to: - **Rely exclusively on the target paper.** The prompt prohibits speculation and external examples, reinforcing adherence to the provided scientific evidence. - **Capture all major findings.** SFT encourages summaries that match the scope of the study, avoiding both omission of central findings and unnecessary broadening. - **Avoid hallucinations.** Exposure to accurate exemplars and strict alignment requirements reduces the likelihood of invented facts or unwarranted claims. **Style.** We expose models to policy-brief writing conventions, including professional tone, academic rigor, and a standardized four-part layout (Policy Problem, Scientific Research Findings, Scientific Research Study Methods, and Policy Implications). Through this structure, the model learns: - **Consistent four-section structure.** The SFT model generates outputs that more reliably follow the required organization. - **Policy-brief writing style.** Fine-tuned outputs exhibit concise, declarative phrasing, policy relevance, and precise causal language. - **Improved higher-level organization.** Transitions between sections become more coherent, with clearer logical connections between findings and implications. - **Reduction of verbosity or filler.** The prompt explicitly disallows padding or oversimplification, teaching the model to maintain informational density without unnecessary elaboration. **Pattern Alignment.** In Table 42, we provide three (paper, brief) examples and a fixed output schema, enabling the model to learn generalizable structural patterns without memorizing specific text. The model acquired: - **Abstract communicative patterns.** These include framing a policy problem, narratively summarizing scientific evidence, and deriving policy implications strictly from the study’s results. - **Cross-paper generalization.** Although each exemplar concerns a different scientific domain, their shared functional structure teaches the model to transfer these patterns to unseen papers. Importantly, the prompt design discourages memorization of surface-level form: - The model does not reproduce specific sentences from the samples. - The vocabulary, phrasing, and order of points vary across target papers, indicating structural rather than textual memorization. **Implicit Skill Acquisition.** Beyond explicit patterns, SFT teaches several latent skills that contribute to improved performance: - **Functional-length adaptation.** The model adjusts the depth of discussion to match the complexity of the target scientific paper, rather than imitating sample length. - **Disciplined scientific reasoning.** The SFT model more consistently avoids speculation, external examples, or causal claims not supported by the paper, and provides accurate interpretations of empirical results. - **Alignment between findings and policy implications.** The fine-tuned model produces implications that more directly follow from the paper’s findings, yielding a clearer and more logically grounded causal chain.
| **Gemma-27B Output** | **Gemma-27B-SFT Output** | **Analysis of Improvements** | |:---|:---|:---| | "Even when rooftop solar is free for low-income households, take-up remains limited because qualified homeowners lack information, face high transaction costs, and may distrust programme providers. Existing referral schemes—which efficiently connect providers with eligible clients—typically rely on financial rewards alone, leaving many referrals unrealised and widening clean-energy equity gaps." | "Subsidized energy assistance programs, like those offering low-income solar installations, are vital for promoting energy justice and an equitable clean energy transition. However, these programs often suffer from low participation rates, leaving eligible households underserved. This under-subscription stems not only from economic barriers but also from non-economic factors like lack of information, complex processes, and distrust, resulting in inefficient allocation of resources and hindering progress towards equitable access to clean energy benefits." | **Factual Grounding + Style:** SFT output maintains stronger factual alignment with the paper’s framing and avoids speculative or generic claims. It exhibits a more formal policy-brief tone ("energy justice", "equitable clean energy transition"), introduces clear causal chains, and avoids compression of complex factors into short sentences. The structure is clearer and more policy-oriented. | | "Adding a small, unconditional US\$1 gift (reciprocity appeal) raised referrals by 2.6-fold compared with a standard reward-only postcard and improved referral quality... Both interventions produced durable effects, continuing to outperform the control nine months after the campaign, indicating that they shifted neither timing nor total referrals." | "Adding a small, non-contingent gift (\$1) alongside a referral reward reminder... and simplifying the referral process... increased referral rates by 2.6 to 5.2 times compared to simply reminding participants of the reward. This translated to 5.2 times more solar contracts at a lower cost per contract... The quality of referrals... remained comparable to the control group, and the increased volume of referrals still resulted in a substantial number of qualified leads." | **Factual Precision:** SFT output reports the quantitative findings more faithfully, reflecting correct magnitudes and their policy relevance (e.g., cost per contract, quality of referrals). **Pattern Alignment:** Follows the narrative structure seen in the training samples—findings grouped by mechanism, then scale, then interpretation. **No Memorization:** Different phrasing and structure from samples; reasoning aligns with the target paper. | | "Researchers partnered with GRID Alternatives to conduct a large, pre-registered randomised controlled trial... Eleven days later, a reminder postcard was sent. Referral behaviour... was tracked for 17 weeks and nine months. Poisson and logistic regressions... were used to estimate treatment effects." | "Researchers partnered with a non-profit administering a low-income solar program and conducted a randomized controlled trial... Data on referral rates, referral quality... and resulting solar contracts were collected... and analyzed using statistical methods including logistic regression and Poisson count models. The study accounted for potential confounding factors through stratification during randomization and robust statistical analysis." | **Implicit Skill Acquisition:** SFT output expands the methods narrative with clearer explanation of data collection, statistical approaches, and rationale—matching the “narrative methods” requirement in the SFT prompt. **Style Enforcement:** Avoids bullet points, uses narrative flow, mirrors sample briefs’ level of technical detail. | | "Pairing financial rewards with reciprocity and simplification multiplies programme reach at minimal incremental cost, making it a practical complement to larger subsidies... Programme managers can tailor interventions to their objectives: reciprocity alone yields higher-quality referrals, while reciprocity + simplification maximises volume and cost-effectiveness." | "Integrate Behavioral Insights... Small Incentives Can Yield Large Returns... Streamline Processes... Targeted Outreach... Cost-Effectiveness... Program Evaluation." | **Style + Structure:** SFT output produces a cleaner, bullet-pointed policy implications section, matching the required format ("bullet points only"). **Pattern Generalization:** SFT consistently aligns implications with findings—no speculative statements, each point tied to an observed effect. **Improved Policy Orientation:** Uses the framing style of the sample briefs (action verbs, policy directives). |
| **DeepSeek-V3 Output** | **Analysis of Limitations** | |:---|:---| | "Climate-tech startups are critical for accelerating the commercialization of innovative technologies needed to meet global climate goals. However, these startups face high failure rates and chronic underinvestment due to long development timelines, high capital requirements, and misaligned investor expectations." | **Factual Grounding:** The model introduces generic and speculative claims (e.g., “chronic underinvestment,” “misaligned investor expectations”) that do not appear in the target paper. It expands beyond the study context, weakening alignment with the evidence. **Style:** Tone resembles an op-ed or general commentary, not a structured policy brief; lacks concise problem framing. **Pattern Alignment:** Does not follow the requested four-section structure and drifts toward thematic discussion rather than evidence-based summary. **Implicit Skills:** Fails to preserve the functional depth or specificity expected from a scientifically grounded brief. | | "Corporate investment is strongly associated with startup success (110% higher exit rate) but also with failure (62% higher failure rate), suggesting corporations selectively fund higher-risk, high-reward ventures." | **Factual Grounding:** The model fabricates numerical results (“110% higher exit rate,” “62% higher failure rate”), which do not appear in the original paper. **Style:** Uses bullet-like enumeration in Findings, deviating from the narrative style required. **Pattern Alignment:** Introduces causal interpretation (“high-risk, high-reward ventures”) not supported by the study, violating evidence-based reporting. **Implicit Skills:** Lacks disciplined reasoning; interprets correlation causally. | | "Applied Cox Proportional Hazards models to assess how funding sources correlate with startup exits and failures." | **Factual Grounding:** Substitutes the study’s actual methods with invented ones (the real paper uses descriptive correlations and regression-based comparisons, not Cox models). **Style:** Overly technical for policy brief norms; deviates from the narrative methods requirement. **Pattern Alignment:** Fails to mirror the structure used in the exemplars, which emphasize high-level methodological narrative rather than model-heavy detail. **Implicit Skills:** Does not adapt the level of technicality appropriately, violating “functional-length” guidance. | | "Target high-risk sectors underserved by private capital (e.g., nuclear, geothermal)... Create platforms... Mitigate risks of private investment... Leverage regional strengths." | **Factual Grounding:** Several points are speculative and not grounded in the paper’s empirical findings (e.g., nuclear or geothermal sectors are not mentioned). **Style:** Bullet points are overly long, mixing policy recommendations unspecified in the study. **Pattern Alignment:** Introduces external examples and normative proposals that diverge from evidence-based implications. **Implicit Skills:** Fails to maintain alignment between Findings → Implications; proposed policies are not traceable back to the study’s results. |
| **GPT-4o Output** | **Analysis of Limitations** | |:---|:---| | "California faces a significant challenge in reducing greenhouse gas (GHG) emissions from its transportation sector, primarily due to high dependence on oil extraction." | **Factual Grounding:** The model reframes the policy problem incorrectly—the paper discusses oil-extraction emissions and community impacts, not transportation-sector dependence. **Style:** Tone is explanatory and generic, not aligned with compact, policy-brief framing. **Pattern Alignment:** Does not follow the structured policy problem template (missing specific grounding in setbacks, excise taxes, and carbon taxes). **Implicit Skills:** Shows poor adherence to the task requirement of rephrasing the study’s actual motivation. | | "Among these, setbacks yield the highest health benefits and equity gains... but also lead to more substantial worker compensation losses compared to excise and carbon taxes." | **Factual Grounding:** Omits key quantitative details and nuances (e.g., study’s nuanced 2045 target scenarios; absence of explicit percentages). **Style:** Narrative is broad rather than specific; lacks explicit causal reasoning tied to the paper’s model results. **Pattern Alignment:** Loses the ordering convention used in exemplars (Findings → quantitative evidence → constraints). **Implicit Skills:** Simplifies complex findings, failing to preserve functional-length or depth appropriate to the study. | | "The study employs a comprehensive methodology combining an empirical oil-production model, an air pollution dispersion model, and an employment input-output model." | **Factual Grounding:** The original paper does not use an input-output model; methodology is partially hallucinated. **Style:** Uses vague terms (“comprehensive methodology”), lacking the structured narrative required in SFT. **Pattern Alignment:** Deviates from exemplar tone, which emphasizes structured description over meta-evaluation. **Implicit Skills:** Does not constrain technical detail to the appropriate depth; misses mention of limitations as done in samples. | | "Policymakers must consider integrating supply-side policies with demand-side strategies to ensure comprehensive and coordinated reductions in statewide and global GHG emissions." | **Factual Grounding:** Introduces policy recommendations not supported by the paper (demand-side strategies are outside the study’s scope). **Style:** Shifts to normative messaging inappropriate for a scientific policy brief. **Pattern Alignment:** Does not adhere to the exemplar requirement that policy implications be strictly derived from the Findings. **Implicit Skills:** Violates “strict factual alignment”—adding novel arguments inconsistent with paper evidence. |
#### Impact of Prompt Length for Tasks 1-4 For Tasks 1-4, we test whether two sentences of context are sufficient for a model to predict the next sentence. Using Task 1 as an example, we evaluate 50 samples with 2-, 3-, and 4-sentence prompts, nesting shorter prompts within the 4-sentence version for fairness. Results in [tab:pro_len_task1] suggest that prompt length has a limited effect on model accuracy. #### Human–LLM Judge Agreement Evaluation In this section, we evaluate whether the LLM-based judge align with human judgments. We conduct two tests to assess agreement between humans and the LLM judge. - **Direct human scoring (13.6.1).** We ask two human evaluators to directly score model outputs and compare their averaged scores with those produced by the Gemini judge. - **Confusion-matrix validation 13.6.2.** We take the top five and bottom five LLaMA-4-Maverick summaries (ranked by Gemini-2.5-Pro on Task 7), have a policy expert label them as “good” or “bad”, and compute a confusion matrix to assess judge reliability. ##### Direct Human Scoring We randomly sample 10 cases from Tasks 9 and 11-15 for human evaluation (using the same 10 cases across Tasks 11–15). Due to the limited number of evaluation questions, we report only the final results and do not conduct bootstrap significance testing. The results show that Gemma-27B-SFT consistently outperforms DeepSeek-V3 and GPT-4o, aligning closely with our reference-based evaluation. For completeness, we present one representative example for Tasks 9 and 15 along with brief analyses in [tab:task9_human,tab:task15_human_part1,tab:task15_human_part2].
##### Confusion-matrix Validation In this part, we evaluate Gemini’s scoring of LLaMA-4-Maverick outputs on Task 7 as a toy example. Specifically, we select the five samples that receive the highest scores and the five that receive the lowest scores from Gemini-2.5-Pro. We then ask our political expert to review these ten samples and determine which are "good" or "bad" summarizations (5 "good" and 5 "bad" summarizations). We show the confusion matrix in 7, and the results show that our reference-free method is indeed reliable.
| **Expert Judgment** | **Predicted Good** | **Predicted Bad** | |:--------------------|:------------------:|:-----------------:| | **Good** | 5 | 0 | | **Bad** | 0 | 5 | **Confusion Matrix of the Gemini-2.5-Pro-based Reference-free Judge.** We evaluate Gemini’s scoring of LLaMA-4-Maverick outputs on Task 7 as a toy example. Specifically, we select the five samples that receive the highest scores and the five that receive the lowest scores from Gemini-2.5-Pro. We then ask our political expert to review these ten samples and determine which are "good" or "bad" summarizations (5 "good" and 5 "bad" summarizations). The confusion matrix shows the reliability of this method.
#### Why Tasks 11-14 Beyond Task 15: Section vs. Full Brief Generation We provide a concrete example comparing LLM output generated section-by-section versus as a full policy brief. Specifically, we use LLaMA-3.1-8B-Instruct on the paper *European battery self-sufficiency* (Link et al. 2025a). We show the results in [tab:sec_vs_full_example] and [tab:sec_vs_full_example_2]. This example illustrates why Tasks 11–14 are necessary in addition to Task 15 in Sci2Pol-Bench. - Task 15 evaluates end-to-end policy brief generation, but Tasks 11-14 aim to capture complementary dimensions of model ability. When LLMs generate briefs one section at a time, they often include more technical details: precise figures, dataset sizes, or methodological constructs. This section-wise mode highlights a model’s capacity to reproduce fine-grained scientific evidence. However, the sections may drift in terminology, exaggerate findings, or fail to connect implications back to results. The outputs are accurate in isolation but lack global coherence. - In contrast, full-brief generation produces more fluent and coherent outputs. The narrative flows smoothly from the policy problem to the findings, methods, and implications, often with organized recommendations and a consistent tone. Yet this approach tends to sacrifice technical precision. Numbers are simplified, mechanisms skipped, and caveats dropped. These lead to overconfident or generic claims that weaken policy relevance. - Thus, Tasks 11-14 complement Task 15 by disentangling precision from coherence. Section-level generation stresses factual grounding and attention to detail, while full-brief generation stresses holistic structure and readability. Evaluating both provides a fuller picture of LLM performance and reveals trade-offs that Task 15 alone cannot capture. These comparisons justify the inclusion of Tasks 11-14 as critical complements to Task 15. #### Over-endorsement Analysis on Tasks 16 and 18 We take Tasks 16 and 18 as examples to conduct an over-endorsement analysis by comparing predicted labels ("SUPPORT" or "CONTRADICT") against ground-truth annotations. Specifically, we compute the number of false positives—instances where the model predicts "SUPPORT" despite the true label being "CONTRADICT"—as well as false negatives, where the model fails to endorse a true "SUPPORT" case by predicting "CONTRADICT". For each task, we calculate the false positive rate (FPR), defined as the proportion of CONTRADICT-labeled instances that are incorrectly predicted as SUPPORT, and the false negative rate (FNR), defined as the proportion of SUPPORT-labeled instances that are incorrectly predicted as CONTRADICT. We also compute the ratio of false positives to false negatives (FPR/FNR ratio). A high FPR/FNR ratio indicates a model’s propensity to over-endorse, potentially overstating the strength of scientific support behind science and policy claims. This analysis provides insight into models’ decision biases, which is critical for assessing their reliability in science and policy-relevant contexts. We show the detailed results in [tab:over_endorse], and we have the following findings: - The FPR/FNR ratios show consistent results across Tasks 16 and 18, with the exception of the model DeepSeek-V3, which exhibits divergent behavior. - Most advanced models do not exhibit over-endorsement, whereas smaller models such as LLaMA-3.3-70B-IT, Gemma-3-27B, Gemma-3-12B, and LLaMA-3.1-8B-IT demonstrate a tendency to over-endorse scientific claims and policy implications. #### Information Leakage Check of In-context Polishing in 3.3 When applying in-context polishing to revise policy briefs (3.3), we use three reference samples from the 85 expert-written briefs. The three reference samples are - *Reliable industrial policies required to support the ramp-up of European battery production* – Paper (Link et al. 2025a); Brief (Link et al. 2025b). - *How central banks address climate and transition risks* – Paper (Shears et al. 2025b); Brief (Shears et al. 2025a). - *Faster deployment of renewables stabilizes electricity prices in Europe* – Paper (Navia Simon and Diaz Anadon 2025b); Brief (Navia Simon and Diaz Anadon 2025a). Although the prompt for GPT-o3 instructs the model to mimic only writing style and format (42), there remains a risk of information leakage. To validate that this step does not introduce leakage, we perform three ablation studies: - Excluding the three samples from Task 15 and then using Task 15 for justification (13.9.1). - Using the three latest published paper-brief pairs (detailed in 15.4) as the in-context examples to polish the dataset, and subsequently using this newly polished dataset for supervised tuning (13.9.2). - Using an additional 8 newly published samples (after May 10th, 2025, detailed in 15.3) as the new samples for Task 15, and testing the model performance on these 8 samples with Task 15 (13.9.3). ##### Excluding Three Samples from Justification We exclude the three samples from Task 15 and use Task 15 for justification. We show the results in [tab:info_leak]. Gains remain stable, and this shows that in-context polishing does not induce leakage. ##### In-Context Polishing with Three New Published Pairs We use the three latest published paper-brief pairs as the three in-context examples to polish the dataset, and use this newly polished dataset to do supervised tuning. The three examples include: - *Nutritional outcomes of irrigation expansion* – Paper (Mehta, Muller, et al. 2025); Brief (Mehta, Müller, et al. 2025). - *Managing development choices is essential to reduce coastal flood risk in China* – Paper (Y. Wang, Ye, Nicholls, Olsson, Vuuren, et al. 2025); Brief (Y. Wang, Ye, Nicholls, Olsson, P. van Vuuren, et al. 2025). - *Heat pumps can help alleviate residential energy insecurity in the USA* – Paper (Ye et al. 2025b); Brief (Ye et al. 2025a). We show the results in [tab:info_leak_a2]. Gains remain stable, and this shows that in-context polishing does not induce leakage. ##### Evaluating on New Eight Test Samples We use another eight newly published paper-brief pairs to generate Task 15. These include all the new peer-reviewed pairs up to Nov. 12th, 2025. We list the details of these eight new pairs in 15.3. We show the results in [tab:info_leak_a3]. Gains remain stable, and this shows that in-context polishing does not induce leakage. #### Athlete as Judge Testing: GPT Family A common concern in benchmark design is potential circularity when GPT family models e.g., GPT-o3) generate intermediate data, such as classification labels or gold prompts for tasks where other GPT family models (e.g., GPT-4o) are later evaluated. This “athlete as judge” setup could favor models from the same family because of shared training signals and writing style. To test this, we design two experiments: - Cross-lineage evaluation: Evaluate the performance of GPT-4o and DeepSeek-V3 on datasets generated by GPT-o3 or a different lineage model, DeepSeek-R1 (13.10.1). - Model substitution: Replace GPT-o3 with MiniMax-M2 in the data construction process (13.10.2). ##### Potential Circularity: GPT vs. DeepSeek Family To test whether this issue affects Sci2Pol-Bench, we conduct a controlled comparison on Task 16. We evaluate GPT-4o and DeepSeek-V3 on datasets generated either by GPT-o3 or by a different lineage model, DeepSeek-R1. As shown in 8, DeepSeek-V3 consistently outperforms GPT-4o, and both models drop 10-12 points when prompts come from DeepSeek-R1 instead of GPT-o3. DeepSeek-R1 prompts are harder. This demonstrates that any benefit from GPT-o3 prompts applies broadly across models rather than giving GPT-family models a special advantage, and it confirms that the “athlete as judge” concern does not affect our setting.
| **Model** | **Dataset (GPT-o3)** | **Dataset (DeepSeek-R1)** | |:----------------|:--------------------:|:-------------------------:| | **GPT-4o** | 98.61 $`\pm`$ 0.39 | 86.93 $`\pm`$ 1.26 | | **DeepSeek-V3** | 98.97 $`\pm`$ 0.37 | 89.00 $`\pm`$ 1.06 | **GPT vs. DeepSeek Family on Task 16.** DeepSeek-V3 consistently outperforms GPT-4o, and both models drop 10-12 points when prompts come from DeepSeek-R1 instead of GPT-o3. DeepSeek-R1 prompts are harder. This demonstrates that any benefit from GPT-o3 prompts applies broadly across models rather than giving GPT family models a special advantage, and it confirms that the “athlete as judge” concern does not affect our setting.
##### Model Substitution by MiniMax-M2 GPT-o3 is utilized to generate data for Tasks 5, 11, 13, 16, and 18, and to polish all corpus documents. To assess the influence of this model choice, we design an ablation study using Tasks 11 and 13 as illustrative examples. The experiment proceeds as follows: (i) We replicate the original data construction method, but replace the data generation model (GPT-o3) with MiniMax-M2. (ii) We then evaluate the performance of GPT-4o, DeepSeek-V3, and Mistral-Large on the data constructed by MiniMax-M2 for Tasks 11 and 13. (iii) The evaluation utilizes a dual LLM Judge setup, employing both Gemini and MiniMax-M2. The full results are presented in [app:tab_replace_gpt_minimax]. These results demonstrate consistency when employing different large language models in the dataset construction process. #### Saturation Testing We do the saturation analysis as in 10. We report the relationship between Sci2Pol-Bench performance and three commonly used model characteristics: (a) model size, (b) pre-training FLOPs, and (c) release date. Although several model developers do not publicly disclose all training details, the approximations used here follow standard practice in the LLM evaluation literature and provide reliable, order-of-magnitude comparisons.
Model Size
Pre-training FLOPs
Release Date
Satuation Testing: Model Size, Pre-training FLOPs, and Release Date.
###### Score vs. Model Size.  Very large models (e.g., GPT-4o) perform strongly, but several mid-sized models such as DeepSeek-V3 match or exceed them, indicating that parameter count alone does not determine policy-reasoning quality. Smaller models ($`<10B`$) consistently underperform, suggesting a minimum capacity threshold. For proprietary models that do not release full architectural details, we use the closest publicly stated or family-level parameter estimate. These values are widely used by open-source leaderboards and preserve the correct relative placement among models. ###### Score vs. Pre-training FLOPs.  Although the highest-FLOPs models generally achieve the strongest results, many models in the $`10^{24}`$–$`10^{25}`$ FLOP range overlap substantially, showing diminishing returns once compute passes a sufficient scale. Several developers no longer publish exact FLOP counts. Therefore, we follow the standard estimation heuristic $`\text{FLOPs} \approx 6 \times N_{\text{params}} \times T_{\text{tokens}}`$ using known or reported model sizes, token counts, and architectural multipliers. Even when exact numbers are unavailable, the order-of-magnitude estimates are robust and do not affect the qualitative scaling trends. ###### Score vs. Release Date.  The third panel visualizes performance relative to the earliest publicly documented release month. While newer models (2024–2025) generally occupy the upper performance range, the significant variation observed within yearly cohorts indicates that recency is not the dominant factor. Instead, our supervised fine-tuning results indicate that alignment strategies and the specific data mixture used in training show a stronger influence on Sci2Pol-Bench performance. Overall, the three panels show that Sci2Pol-Bench performance does not follow a simple scaling law. Model size, compute, and release date each matter, but do not fully explain cross-model differences. Instead, high performance is often associated with improved alignment strategies, targeted supervised finetuning, and richer training mixtures. These are factors that help models reason about scientific evidence and policy trade-offs beyond what raw scale alone can provide. #### Comparison of Writing from an Abstract, an Introduction, and a Full Paper We compare policy brief generation when the input is the abstract, the introduction, or the full scientific paper. Writing from the full paper requires a much longer input context. This increases inference cost, fine-tuning cost, and technical difficulty. At the same time, it offers richer information for grounding policy-relevant insights. We first present a concrete example using Grok-3-beta to generate policy briefs for the paper *How central banks address climate and transition risks* (Shears et al. 2025b), based on three input scopes: (i) the full paper, (ii) the introduction only, and (iii) the abstract only. We show the generated outputs in [tab:task16_aligned] and [tab:task16_aligned2]. The outputs enable a detailed comparison across input settings as follows. - When a model writes from the abstract only, it captures the headline claim, the main outcome, and the high-level implications. The tone reads confident, and the structure feels clean. However, the draft lacks mechanism, methods detail, and scope limits. Numbers, when present, drift toward round values, and uncertainty language fades. Policy advice turns generic, because the draft does not anchor actions to specific evidence or design choices. The abstract frames motivation and result, but it does not carry the caveats that make a brief credible, so the model tends to overstate precision and under-specify risks. - When a model writes from the introduction only, it provides better context. The draft states the policy stakes, defines terms, and maps the debate. The flow improves, and the problem framing reads stronger. Yet, the introduction centers motivation and gaps, not verification. Methods remain thin, and heterogeneity across cases stays offstage. Claims may adopt the paper’s persuasive setup, which can tilt the brief toward advocacy language. Without the full evidence base, the draft still softens uncertainty, skips mechanism, and treats recommendations at the slogan level, not at the instrument level. - When a model writes from the full paper, it can link evidence to tools. The draft cites methods, describes mechanisms, and carries uncertainty statements. It notes limits, and it distinguishes what is feasible from what is likely. The narrative supports actionable advice, because the draft can name concrete instruments, timing, and trade-offs. Here, the main risks shift. The model can drown readers in detail, mix minor numbers into the main thread, or copy framing that belongs in a technical appendix. If the paper is dense, the model may misread tables or blend results, so careful summarization still matters.
### Cross-Domain Generalization Testing We test the model’s generalization capacity across domains outside of its primary training focus. Due to the scarcity of readily available peer-reviewed scientific paper-policy brief pairs in other domains, we conduct an exploratory evaluation using one representative paper from each of the following important fields: AI, Robotics, Mathematics, Economics, and Humanities. The generated policy briefs are then subjected to human evaluation. We test three models: Gemma-27B-SFT, DeepSeek-V3, and GPT-4o. The resulting evaluation indicates that our supervised fine-tuning successfully generalizes to other domains. We provide the details of 5 papers as follows: - AI: “Generative AI needs adaptive governance (Reuel and Undheim 2024)”. We show the results in [app:tab_ai_1,app:tab_ai_2] - Robotics: “Anticipatory policy design for robots in public spaces (Mintrom et al. 2025)”. We show the results in [app:tab_rob_1,app:tab_rob_2] - Mathematics: “Educational policy to improve mathematics instruction at scale: Conceptualizing contextual factors (Ryve and Hemmi 2019)”. We show the results in [app:tab_math_1,app:tab_math_2]. - Economics: “Financial inclusion, the shadow economy and economic growth in developing economies (Younas et al. 2022)”. We show the results in [app:tab_econ_1,app:tab_econ_2]. - Humanities: “Science cited in policy documents: Evidence from the Overton database (Fang et al. 2024)”. We show the results in [app:tab_hum_1,app:tab_hum_2].
### Details of 85 Expert-written Paper-brief Pairs We provide the detailed list of 85 pairs in 15.1 and their yearly distribution in 15.2. We also provide the detailed list of 3 new pairs for in-context polishing in 15.4, and 8 new pairs for Task 15 in 15.3. #### List of 85 Expert-written Paper-brief Pairs We document the 85 expert-written scientific paper-policy brief pairs included in Sci2Pol-Bench, citing each source to ensure transparency and reproducibility. 1. Paper (Link et al. 2025a) (DOI: 10.1038/s41560-025-01722-y); Brief (Link et al. 2025b) (DOI: 10.1038/s41560-025-01741-9); Nature Energy; Accessed: 2025-05-10. 2. Paper (Shears et al. 2025b) (DOI: 10.1038/s41560-025-01724-w); Brief (Shears et al. 2025a) (DOI: 10.1038/s41560-025-01725-9); Nature Energy; Accessed: 2025-05-10. 3. Paper (Navia Simon and Diaz Anadon 2025b) (DOI: 10.1038/s41560-025-01704-0); Brief (Navia Simon and Diaz Anadon 2025a) (DOI: 10.1038/s41560-025-01715-x); Nature Energy; Accessed: 2025-05-10. 4. Paper (Heerden et al. 2025b) (DOI: 10.1038/s41560-025-01703-1); Brief (Heerden et al. 2025a) (DOI: 10.1038/s41560-025-01721-z); Nature Energy; Accessed: 2025-05-10. 5. Paper (Millinger, Fredrik Hedenus, et al. 2025) (DOI: 10.1038/s41560-024-01693-6); Brief (Millinger, F. Hedenus, et al. 2025) (DOI: 10.1038/s41560-024-01685-6); Nature Energy; Accessed: 2025-05-10. 6. Paper (Odenweller and Ueckerdt 2025b) (DOI: 10.1038/s41560-024-01684-7); Brief (Odenweller and Ueckerdt 2025a) (DOI: 10.1038/s41560-024-01682-9); Nature Energy; Accessed: 2025-05-10. 7. Paper (Caggiano et al. 2024b) (DOI: 10.1038/s41560-024-01603-w); Brief (Caggiano et al. 2024a) (DOI: 10.1038/s41560-024-01585-9); Nature Energy; Accessed: 2025-05-10. 8. Paper (O’Shaughnessy et al. 2024b) (DOI: 10.1038/s41560-024-01546-2); Brief (O’Shaughnessy et al. 2024a) (DOI: 10.1038/s41560-024-01575-x); Nature Energy; Accessed: 2025-05-10. 9. Paper (Kennedy et al. 2024b) (DOI: 10.1038/s41560-024-01530-w); Brief (Kennedy et al. 2024a) (DOI: 10.1038/s41560-024-01554-2); Nature Energy; Accessed: 2025-05-10. 10. Paper (Sitarz et al. 2024a) (DOI: 10.1038/s41560-024-01505-x); Brief (Sitarz et al. 2024b) (DOI: 10.1038/s41560-024-01545-3); Nature Energy; Accessed: 2025-05-10. 11. Paper (Link et al. 2024b) (DOI: 10.1038/s41560-024-01531-9); Brief (Link et al. 2024a) (DOI: 10.1038/s41560-024-01555-1); Nature Energy; Accessed: 2025-05-10. 12. Paper (White et al. 2024a) (DOI: 10.1038/s41560-023-01422-5); Brief (White et al. 2024b) (DOI: 10.1038/s41560-023-01433-2); Nature Energy; Accessed: 2025-05-10. 13. Paper (Wolske et al. 2023b) (DOI: 10.1038/s41560-023-01298-5); Brief (Wolske et al. 2023a) (DOI: 10.1038/s41560-023-01323-7); Nature Energy; Accessed: 2025-05-10. 14. Paper (Deshmukh et al. 2023a) (DOI: 10.1038/s41560-023-01259-y); Brief (Deshmukh et al. 2023b) (DOI: 10.1038/s41560-023-01273-0); Nature Energy; Accessed: 2025-05-10. 15. Paper (Gars et al. 2022b) (DOI: 10.1038/s41560-022-01122-6); Brief (Gars et al. 2022a) (DOI: 10.1038/s41560-022-01125-3); Nature Energy; Accessed: 2025-05-10. 16. Paper (Gruber et al. 2022a) (DOI: 10.1038/s41560-022-00994-y); Brief (Gruber et al. 2022b) (DOI: 10.1038/s41560-022-01031-8); Nature Energy; Accessed: 2025-05-10. 17. Paper (Longden, Quilty, Riley, White, Klerck, Vanessa Napaltjari Davis, et al. 2022) (DOI: 10.1038/s41560-021-00942-2); Brief (Longden, Quilty, Riley, White, Klerck, Vanessa Napaltjarri Davis, et al. 2022) (DOI: 10.1038/s41560-021-00968-6); Nature Energy; Accessed: 2025-05-10. 18. Paper (Pachauri et al. 2021a) (DOI: 10.1038/s41560-021-00911-9); Brief (Pachauri et al. 2021b) (DOI: 10.1038/s41560-021-00939-x); Nature Energy; Accessed: 2025-05-10. 19. Paper (Hall et al. 2021b) (DOI: 10.1038/s41560-021-00781-1); Brief (Hall et al. 2021a) (DOI: 10.1038/s41560-021-00821-w); Nature Energy; Accessed: 2025-05-10. 20. Paper (Kitzing et al. 2020a) (DOI: 10.1038/s41560-020-00717-1); Brief (Kitzing et al. 2020b) (DOI: 10.1038/s41560-020-00733-1); Nature Energy; Accessed: 2025-05-10. 21. Paper (Bonan et al. 2020b) (DOI: 10.1038/s41560-020-00719-z); Brief (Bonan et al. 2020a) (DOI: 10.1038/s41560-020-00727-z); Nature Energy; Accessed: 2025-05-10. 22. Paper (Goldstein et al. 2020a) (DOI: 10.1038/s41560-020-00683-8); Brief (Goldstein et al. 2020b) (DOI: 10.1038/s41560-020-00691-8); Nature Energy; Accessed: 2025-05-10. 23. Paper (Mani et al. 2020b) (DOI: 10.1038/s41560-020-0596-7); Brief (Mani et al. 2020a) (DOI: 10.1038/s41560-020-0635-4); Nature Energy; Accessed: 2025-05-10. 24. Paper (Casey et al. 2020b) (DOI: 10.1038/s41560-020-0600-2); Brief (Casey et al. 2020a) (DOI: 10.1038/s41560-020-0622-9); Nature Energy; Accessed: 2025-05-10. 25. Paper (Kontokosta et al. 2020b) (DOI: 10.1038/s41560-020-0589-6); Brief (Kontokosta et al. 2020a) (DOI: 10.1038/s41560-020-0603-z); Nature Energy; Accessed: 2025-05-10. 26. Paper (Kaufmann and Connelly 2020b) (DOI: 10.1038/s41560-020-0549-1); Brief (Kaufmann and Connelly 2020a) (DOI: 10.1038/s41560-020-0563-3); Nature Energy; Accessed: 2025-05-10. 27. Paper (Braunholtz-Speight et al. 2020a) (DOI: 10.1038/s41560-019-0546-4); Brief (Braunholtz-Speight et al. 2020b) (DOI: 10.1038/s41560-020-0556-2); Nature Energy; Accessed: 2025-05-10. 28. Paper (Kar et al. 2019) (DOI: 10.1038/s41560-019-0429-8); Brief (Kar et al. 2020) (DOI: 10.1038/s41560-019-0536-6); Nature Energy; Accessed: 2025-05-10. 29. Paper (Liu and Rajagopal 2019) (DOI: 10.1038/s41560-019-0430-2); Brief (Rajagopal and Liu 2020) (DOI: 10.1038/s41560-019-0532-x); Nature Energy; Accessed: 2025-05-10. 30. Paper (White and Sintov 2020a) (DOI: 10.1038/s41560-019-0507-y); Brief (White and Sintov 2020b) (DOI: 10.1038/s41560-019-0515-y); Nature Energy; Accessed: 2025-05-10. 31. Paper (Rinscheid and Wüstenhagen 2019b) (DOI: 10.1038/s41560-019-0460-9); Brief (Rinscheid and Wüstenhagen 2019a) (DOI: 10.1038/s41560-019-0509-9); Nature Energy; Accessed: 2025-05-10. 32. Paper (Mays et al. 2019a) (DOI: 10.1038/s41560-019-0476-1); Brief (Mays et al. 2019b) (DOI: 10.1038/s41560-019-0502-3); Nature Energy; Accessed: 2025-05-10. 33. Paper (Egli et al. 2018) (DOI: 10.1038/s41560-018-0277-y); Brief (Egli et al. 2019) (DOI: 10.1038/s41560-019-0482-3); Nature Energy; Accessed: 2025-05-10. 34. Paper (Tiefenbeck et al. 2019a) (DOI: 10.1038/s41560-018-0282-1); Brief (Tiefenbeck et al. 2019b) (DOI: 10.1038/s41560-019-0480-5); Nature Energy; Accessed: 2025-05-10. 35. Paper (Azarova et al. 2018) (DOI: 10.1038/s41560-018-0105-4); Brief (Azarova et al. 2019) (DOI: 10.1038/s41560-019-0479-y); Nature Energy; Accessed: 2025-05-10. 36. Paper (Apostoleris et al. 2018) (DOI: 10.1038/s41560-018-0256-3); Brief (Apostoleris et al. 2019) (DOI: 10.1038/s41560-019-0481-4); Nature Energy; Accessed: 2025-05-10. 37. Paper (Mahdavi et al. 2025a) (DOI: 10.1038/s41558-025-02283-4); Brief (Mahdavi et al. 2025b) (DOI: 10.1038/s41558-025-02304-2); Nature Climate Change; Accessed: 2025-05-10. 38. Paper (Ogier et al. 2025b) (DOI: 10.1038/s41558-025-02291-4); Brief (Ogier et al. 2025a) (DOI: 10.1038/s41558-025-02292-3); Nature Climate Change; Accessed: 2025-05-10. 39. Paper (Y. Liu et al. 2025a) (DOI: 10.1038/s41558-024-02237-2); Brief (Y. Liu et al. 2025b) (DOI: 10.1038/s41558-024-02240-7); Nature Climate Change; Accessed: 2025-05-10. 40. Paper (Tang et al. 2024a) (DOI: 10.1038/s41558-024-02162-4); Brief (Tang et al. 2024b) (DOI: 10.1038/s41558-024-02145-5); Nature Climate Change; Accessed: 2025-05-10. 41. Paper (Druckenmiller et al. 2024a) (DOI: 10.1038/s41558-024-02082-3); Brief (Druckenmiller et al. 2024b) (DOI: 10.1038/s41558-024-02083-2); Nature Climate Change; Accessed: 2025-05-10. 42. Paper (Nowak et al. 2024b) (DOI: 10.1038/s41558-024-02054-7); Brief (Nowak et al. 2024a) (DOI: 10.1038/s41558-024-02055-6); Nature Climate Change; Accessed: 2025-05-10. 43. Paper (Lamb et al. 2024b) (DOI: 10.1038/s41558-024-01984-6); Brief (Lamb et al. 2024a) (DOI: 10.1038/s41558-024-01993-5); Nature Climate Change; Accessed: 2025-05-10. 44. Paper (Gasparini, M. C. Ives, et al. 2024) (DOI: 10.1038/s41558-024-01972-w); Brief (Gasparini, M. Ives, et al. 2024) (DOI: 10.1038/s41558-024-01959-7); Nature Climate Change; Accessed: 2025-05-10. 45. Paper (Duan et al. 2024a) (DOI: 10.1038/s41558-024-01952-0); Brief (Duan et al. 2024b) (DOI: 10.1038/s41558-024-01962-y); Nature Climate Change; Accessed: 2025-05-10. 46. Paper (Andreoni et al. 2024b) (DOI: 10.1038/s41558-023-01870-7); Brief (Andreoni et al. 2024a) (DOI: 10.1038/s41558-023-01871-6); Nature Climate Change; Accessed: 2025-05-10. 47. Paper (Linsenmeier et al. 2023a) (DOI: 10.1038/s41558-023-01710-8); Brief (Linsenmeier et al. 2023b) (DOI: 10.1038/s41558-023-01700-w); Nature Climate Change; Accessed: 2025-05-10. 48. Paper (Merfort et al. 2023a) (DOI: 10.1038/s41558-023-01697-2); Brief (Merfort et al. 2023b) (DOI: 10.1038/s41558-023-01711-7); Nature Climate Change; Accessed: 2025-05-10. 49. Paper (Cerf et al. 2023b) (DOI: 10.1038/s41558-023-01679-4); Brief (Cerf et al. 2023a) (DOI: 10.1038/s41558-023-01677-6); Nature Climate Change; Accessed: 2025-05-10. 50. Paper (Buck et al. 2023b) (DOI: 10.1038/s41558-022-01592-2); Brief (Buck et al. 2023a) (DOI: 10.1038/s41558-023-01614-7); Nature Climate Change; Accessed: 2025-05-10. 51. Paper (Harring et al. 2023a) (DOI: 10.1038/s41558-023-01597-5); Brief (Harring et al. 2023b) (DOI: 10.1038/s41558-023-01609-4); Nature Climate Change; Accessed: 2025-05-10. 52. Paper (Basheer et al. 2023a) (DOI: 10.1038/s41558-022-01556-6); Brief (Basheer et al. 2023b) (DOI: 10.1038/s41558-022-01557-5); Nature Climate Change; Accessed: 2025-05-10. 53. Paper (Iyer et al. 2022a) (DOI: 10.1038/s41558-022-01508-0); Brief (Iyer et al. 2022b) (DOI: 10.1038/s41558-022-01517-z); Nature Climate Change; Accessed: 2025-05-10. 54. Paper (Ruig et al. 2022b) (DOI: 10.1038/s41558-022-01501-7); Brief (Ruig et al. 2022a) (DOI: 10.1038/s41558-022-01502-6); Nature Climate Change; Accessed: 2025-05-10. 55. Paper (Bjørn et al. 2022b) (DOI: 10.1038/s41558-022-01379-5); Brief (Bjørn et al. 2022a) (DOI: 10.1038/s41558-022-01385-7); Nature Climate Change; Accessed: 2025-05-10. 56. Paper (Mildenberger et al. 2022b) (DOI: 10.1038/s41558-021-01268-3); Brief (Mildenberger et al. 2022a) (DOI: 10.1038/s41558-021-01270-9); Nature Climate Change; Accessed: 2025-05-10. 57. Paper (Budolfson et al. 2021a) (DOI: 10.1038/s41558-021-01217-0); Brief (Budolfson et al. 2021b) (DOI: 10.1038/s41558-021-01228-x); Nature Climate Change; Accessed: 2025-05-10. 58. Paper (Janssens et al. 2020) (DOI: 10.1038/s41558-020-0847-4); Brief (Janssens et al. 2021) (DOI: 10.1038/s41558-021-01201-8); Nature Climate Change; Accessed: 2025-05-10. 59. Paper (Moffette et al. 2021b) (DOI: 10.1038/s41558-020-00956-w); Brief (Moffette et al. 2021a) (DOI: 10.1038/s41558-021-01195-3); Nature Climate Change; Accessed: 2025-05-10. 60. Paper (Bechtel et al. 2020) (DOI: 10.1038/s41558-020-00914-6); Brief (Bechtel et al. 2021) (DOI: 10.1038/s41558-021-01202-7); Nature Climate Change; Accessed: 2025-05-10. 61. Paper (Peng et al. 2021a) (DOI: 10.1038/s41558-021-01128-0); Brief (Peng et al. 2021b) (DOI: 10.1038/s41558-021-01193-5); Nature Climate Change; Accessed: 2025-05-10. 62. Paper (Tran et al. 2024) (DOI: 10.1038/s44284-024-00116-7); Brief (Ivanov et al. 2024) (DOI: 10.1038/s44284-024-00128-3); Nature Cities; Accessed: 2025-05-10. 63. Paper (Diezmartı́nez et al. 2024b) (DOI: 10.1038/s44284-024-00121-w); Brief (Diezmartı́nez et al. 2024a) (DOI: 10.1038/s44284-024-00129-2); Nature Cities; Accessed: 2025-05-10. 64. Paper (Mollborn et al. 2025a) (DOI: 10.1177/00221465241255946); Brief (Mollborn et al. 2025b) (DOI: 10.1177/00221465251315281); Journal Of Health And Social Behavior; Accessed: 2025-05-10. 65. Paper (Moinester and Stanhope 2024a) (DOI: 0.1177/00221465241230839); Brief (Moinester and Stanhope 2024b) (DOI: 10.1177/00221465241269117); Journal Of Health And Social Behavior; Accessed: 2025-05-10. 66. Paper (Han et al. 2024a) (DOI: 10.1177/00221465231205266); Brief (Han et al. 2024b) (DOI: 10.1177/00221465241248972); Journal Of Health And Social Behavior; Accessed: 2025-05-10. 67. Paper (Dore et al. 2024a) (DOI: 10.1177/00221465231194043); Brief (Dore et al. 2024b) (DOI: 10.1177/00221465241226808); Journal Of Health And Social Behavior; Accessed: 2025-05-10. 68. Paper (McFarland et al. 2023a) (DOI: 10.1177/00221465221109202); Brief (McFarland et al. 2023b) (DOI: 10.1177/00221465221150307); Journal Of Health And Social Behavior; Accessed: 2025-05-10. 69. Paper (Masters et al. 2023a) (DOI: 10.1177/00221465231165284); Brief (Masters et al. 2023b) (DOI: 10.1177/00221465231171627); Journal Of Health And Social Behavior; Accessed: 2025-05-10. 70. Paper (Parbst and Wheaton 2023a) (DOI: 10.1177/00221465231166334); Brief (Parbst and Wheaton 2023b) (DOI: 10.1177/00221465231190977); Journal Of Health And Social Behavior; Accessed: 2025-05-10. 71. Paper (Czarnecki et al. 2023a) (DOI: 10.1177/00221465231172177); Brief (Czarnecki et al. 2023b) (DOI: 10.1177/00221465231209380); Journal Of Health And Social Behavior; Accessed: 2025-05-10. 72. Paper (McCabe 2022a) (DOI: 10.1177/00221465211058152); Brief (McCabe 2022b) (DOI: 10.1177/00221465221097453); Journal Of Health And Social Behavior; Accessed: 2025-05-10. 73. Paper (Vuolo et al. 2022a) (DOI: 10.1177/00221465211067209); Brief (Vuolo et al. 2022b) (DOI: 10.1177/00221465221112986); Journal Of Health And Social Behavior; Accessed: 2025-05-10. 74. Paper (Anderson and Ray-Warren 2022a) (DOI: 10.1177/00221465221074915); Brief (Anderson and Ray-Warren 2022b) (DOI: 10.1177/00221465221130917); Journal Of Health And Social Behavior; Accessed: 2025-05-10. 75. Paper (Augustine 2021a) (DOI: 10.1177/0022146520979664); Brief (Augustine 2021b) (DOI: 10.1177/0022146520986008); Journal Of Health And Social Behavior; Accessed: 2025-05-10. 76. Paper (Manzer and Bell 2021a) (DOI: 10.1177/00221465211003232); Brief (Manzer and Bell 2021b) (DOI: 10.1177/00221465211008328); Journal Of Health And Social Behavior; Accessed: 2025-05-10. 77. Paper (Berg et al. 2021a) (DOI: 10.1177/00221465211052568); Brief (Berg et al. 2021b) (DOI: 10.1177/00221465211055925); Journal Of Health And Social Behavior; Accessed: 2025-05-10. 78. Paper (Schnittker and Do 2020a) (DOI: 10.1177/0022146519899115); Brief (Schnittker and Do 2020b) (DOI: 10.1177/0022146520903969); Journal Of Health And Social Behavior; Accessed: 2025-05-10. 79. Paper (Owens 2020a) (DOI: 10.1177/0022146520924810); Brief (Owens 2020b) (DOI: 10.1177/0022146520926100); Journal Of Health And Social Behavior; Accessed: 2025-05-10. 80. Paper (Thombs et al. 2020a) (DOI: 10.1177/0022146520939514); Brief (Thombs et al. 2020b) (DOI: 10.1177/0022146520945607); Journal Of Health And Social Behavior; Accessed: 2025-05-10. 81. Paper (Bierman and Schieman 2020a) (DOI: 10.1177/0022146520970190); Brief (Bierman and Schieman 2020b) (DOI: 10.1177/0022146520968770); Journal Of Health And Social Behavior; Accessed: 2025-05-10. 82. Paper (Zhang et al. 2025b) (DOI: 10.1038/s41893-025-01533-9); Brief (Zhang et al. 2025a) (DOI: 10.1038/s41893-025-01535-7); Nature Sustainability; Accessed: 2025-05-10. 83. Paper (Azar 2024a) (DOI: 10.1177/00221465241265435); Brief (Azar 2024b) (DOI: 10.1177/00221465241291690); Journal Of Health And Social Behavior; Accessed: 2025-05-10. 84. Paper (Rapp et al. 2022a) (DOI: 10.1177/00221465211058153); Brief (Rapp et al. 2022b) (DOI: 10.1177/00221465211073836); Journal Of Health And Social Behavior; Accessed: 2025-05-10. 85. Paper (Ci et al. 2024b) (DOI: 10.1038/s41893-024-01471-y); Brief (Ci et al. 2024a) (DOI: 10.1038/s41893-024-01472-x); Nature Sustainability; Accessed: 2025-05-10. #### Publication Year Distribution of 85 Policy Briefs As shown in 9, the number of policy briefs grows steadily. This trend highlights the increasing availability of high-quality published data. This suggests strong potential for expanding Sci2Pol-Bench and Sci2Pol-Corpus in the future.
| **Year** | 2019 | 2020 | 2021 | 2022 | 2023 | 2024 | 2025 | |:------------------|:----:|:----:|:----:|:----:|:----:|:----:|:----:| | **Num. of Pairs** | 6 | 15 | 10 | 11 | 12 | 20 | 11 | **Publication Year Distribution of the 85 Expert-written Policy Briefs.**
#### List of 8 New Expert-written Paper-brief Pairs We document the new 8 expert-written scientific paper-policy brief pairs for Task 15. 1. Paper (Weber et al. 2025b) (DOI: 10.1038/s41893-025-01676-9); Brief (Weber et al. 2025a) (DOI: 10.1038/s41893-025-01680-z); Nature Sustainability; Accessed: 2025-05-10. 2. Paper (Minten et al. 2025b) (DOI: 10.1038/s41893-025-01618-5); Brief (Minten et al. 2025a) (DOI: 10.1038/s41893-025-01622-9); Nature Sustainability; Accessed: 2025-05-10. 3. Paper (Vivier et al. 2025b) (DOI: 10.1038/s41558-025-02348-4); Brief (Vivier et al. 2025a) (DOI: 10.1038/s41558-025-02342-w); Nature Climate Change; Accessed: 2025-05-10. 4. Paper (Cai et al. 2025) (DOI: 10.1038/s41893-025-01560-6); Brief (X. Wang et al. 2025) (DOI: 10.1038/s41893-025-01582-0); Nature Sustainability; Accessed: 2025-05-10. 5. Paper (Hermansen et al. 2025a) (DOI: 10.1038/s41586-025-09259-6); Brief (Hermansen et al. 2025b) (DOI: 10.1038/d41586-025-02546-2); Nature; Accessed: 2025-05-10. 6. Paper (Sun et al. 2025b) (DOI: 10.1038/s41560-025-01821-w); Brief (Sun et al. 2025a) (DOI: 10.1038/s41560-025-01822-9); Nature Energy; Accessed: 2025-05-10. 7. Paper (Anshassi and Townsend 2025a) (DOI: 10.1038/s41893-025-01607-8); Brief (Anshassi and Townsend 2025b) (DOI: 10.1038/s41893-025-01597-7); Nature Sustainability; Accessed: 2025-05-10. 8. Paper (Dipoppa and Gulzar 2024) (DOI: 10.1038/s41586-024-08046-z); Brief (Dipoppa and Gulzar 2025) (DOI: 10.1038/d41586-025-02545-3); Nature; Accessed: 2025-05-10. #### List of 3 New In-context Expert-written Paper-brief Pairs We document the new 3 expert-written scientific paper-policy brief pairs for in-context polishing. 1. Paper (Mehta, Muller, et al. 2025) (DOI: 10.1038/s41893-025-01584-y); Brief (Mehta, Müller, et al. 2025) (DOI: 10.1038/s41893-025-01593-x); Nature Sustainability; Accessed: 2025-05-10. 2. Paper (Y. Wang, Ye, Nicholls, Olsson, Vuuren, et al. 2025) (DOI: 10.1038/s41558-025-02439-2); Brief (Y. Wang, Ye, Nicholls, Olsson, P. van Vuuren, et al. 2025) (DOI: 10.1038/s41558-025-02418-7); Nature Climate Change; Accessed: 2025-05-10. 3. Paper (Ye et al. 2025b) (DOI: 10.1038/s41560-025-01845-2); Brief (Ye et al. 2025a) (DOI: 10.1038/s41560-025-01840-7); Nature Energy; Accessed: 2025-05-10. ### Examples for Tasks 1-18 In this section, we provide examples for Tasks 1-18, as show in [tab:exp_task1,tab:exp_task2,tab:exp_task3,tab:exp_task4,tab:exp_task5,tab:exp_task6,tab:exp_task7,tab:exp_task8,tab:exp_task9,tab:exp_task10,tab:exp_task12,tab:exp_task13,tab:exp_task14,tab:exp_task15,tab:exp_task16,tab:exp_task17,tab:exp_task18,tab:exp_task19].
Example for Task 1: Scientific Text Autocompletion.
Prompt: You are given the start of a paragraph from a scientific research paper, ending in “...”. Choose the most coherent continuation from the options below.

Query:

We do not find an association between exposure to climate hazards and re-risking. Instead, we find that de-risking is positively and significantly correlated with higher exposure to physical climate risks. ...

Options:

A. This is puzzling and requires further analysis.

B. Similarly, a large financial sector increases the direct exposure of an economy to stranded asset risks, which could threaten financial stability.

C. It follows that if a country is highly exposed to physical climate risks, it would adopt these practices.

D. In fact, prior research has demonstrated a correlation between physical risks and central bank management of climate risks.

E. We measure climate hazard exposure by using the exposure component of the Notre Dame Global Adaptation Initiative (ND-Gain) country index.
Answer with the letter (A-E) corresponding to the best continuation. Strictly follow this format–do not include any explanations or additional text.
Answer: A
| | |:---| | **Prompt:** | | You are given the start of a paragraph from a policy brief, ending in “...”. Choose the most coherent continuation from the options below. | | | | Query: | | The rapid diffusion of battery electric vehicles, in addition to the decarbonization of the energy sector—requires an increasing number of batteries. However, the EU’s goal to cover 90% of its battery demand from domestic production by 2030 is at risk, as projected demand will likely exceed 1.0 TWh per year and outpace production capacity despite highly ambitious growth rates. ... | | | | Options: | | A. If Europe fails to scale up production, it may face severe economic and geopolitical risks, due to increased dependence on external suppliers, weakened industrial competitiveness and potential for delayed decarbonization. | | B. An urgent question is therefore whether Europe can realistically meet its future battery demand through domestic production, and what policy actions are needed to ensure success. | | C. We find that European battery cell demand will likely surpass 1.0 TWh per year by 2030, whereas domestic production capacity is expected to fall short, creating a risk of supply constraints. | | D. Although Europe can be expected to meet at least 50-60% of its demand through domestic production by 2030, achieving the EU’s 90% self-sufficiency target is feasible but uncertain, as nearly half of our modelled scenarios fail to meet this target (Fig. 1). | | E. If Europe wants more independence from battery cell imports, our findings highlight the urgency of accelerating production capacity expansion, scaling up a battery supply chain, and implementing strong industrial policies to support competitiveness and supply sovereignty. | | | | Answer with the letter (A-E) corresponding to the best continuation. **Strictly follow this format–do not include any explanations or additional text.** | | **Answer:** B | **Example for Task 2: Political Text Autocompletion.**
| | |:---| | **Prompt:** | | You are given three shuffled sentences that originally formed a coherent paragraph from a scientific research paper. Your task is to determine the correct order by selecting the most logical and coherent sequence. | | | | Shuffled Sentences: | | A. Other key markets, such as the United States and China, have also set ambitious ZEV targets from the 2030s. | | B. Accordingly, BEVs prevail in the future portfolios of car manufacturers and several European countries will enforce 100% zero-emission vehicle (ZEV) sales for cars by at least 2035, banning large-scale sales of conventional vehicles as sufficient quantities of sustainable fuels are unlikely. | | C. While some studies have emphasized the difficulties involved in decarbonizing transport, there is robust evidence that battery electric vehicles (BEVs) will form the backbone of future low-carbon road transport. | | | | Answer with a permutation of A, B, and C that best restores the original paragraph (e.g., BAC, CAB). **Strictly follow this format–do not include any explanations or additional text.** | | **Answer:** CBA | **Example for Task 3: Scientific Sentence Reordering.**
| | |:---| | **Prompt:** | | You are given three shuffled sentences that originally formed a coherent paragraph from a policy brief document. Your task is to determine the correct order by selecting the most logical and coherent sequence. | | | | Shuffled Sentences: | | A. Our approach accounts for uncertain ties such as construction delays, utilization rates, and evolving market conditions, and assesses corresponding raw material needs. | | B. Our study uses probabilistic modelling to project future battery demand and domestic production in Europe and evaluates Europe’s pathway towards battery self-sufficiency via probabilistic statements. | | C. We independently model demand and supply using S-shaped diffusion curves based on historical data, actual announced production capacities, and practice-oriented findings about how these announced capacities materialize over time. | | | | Answer with a permutation of A, B, and C that best restores the original paragraph (e.g., BAC, CAB). **Strictly follow this format–do not include any explanations or additional text.** | | **Answer:** CAB | **Example for Task 4: Political Sentence Reordering.**
| | |:---| | **Prompt:** | | You are given a sentence or paragraph from a scientific research paper. Classify the text into one of the following five categories relevant to policy briefs: | | | | 1\. Policy Problem - Describes the societal or strategic issue that the research aims to address. | | 2\. Scientific Research Findings - Reports key empirical results, model outcomes, or discoveries from the study. | | 3\. Scientific Research Study Methods - Details the data sources, methodologies, models, or analytical procedures used. | | 4\. Policy Implications - Explains how the research findings can inform, influence, or support policy decisions. | | 5\. None - Does not fit into any of the above categories. | | | | Your response must be exactly one of the following: | | Policy Problem Scientific Research Findings Scientific Research Study Methods Policy Implications None | | | | Text: | | Batteries are critical to mitigate global warming, with battery electric vehicles as the backbone of low-carbon transport and the main driver of advances and demand for battery technology. However, the future demand and production of batteries remain uncertain, while the ambition to strengthen national capabilities and self-sufficiency is gaining momentum. In this study, leveraging probabilistic modelling, we assessed Europe’s capability to meet its future demand for high-energy batteries via domestic cell production. We found that demand in Europe is likely to exceed 1.0 TWh yr-1 by 2030 and thereby outpace domestic production, with production required to grow at highly ambitious growth rates of 31-68% yr-1. European production is very likely to cover at least 50-60% of the domestic demand by 2030, while 90% self-sufficiency seems feasible but far from certain. Thus, domestic production shortfalls are more likely than not. | | | | Category: | | **Answer:** Policy Problem | **Example for Task 5: Sentence Classification.**
| | |:---| | **Prompt:** | | You are given a multiple-choice question. Read the question carefully and select the best answer from the provided options. Each option is labeled with a capital letter (A, B, C, ...). Only one answer is correct. Provide your response as a single letter (e.g., A, B, C, ...). | | | | Question: | | An ileostomy opening should be: | | | | Options: | | A. less than 1cm in length. | | B. invisible to the naked eye. | | C. 3-5cm in length. | | D. below skin level. | | E. protruding 10cm from the skin. | | F. over 10cm in length. | | G. over 5cm in length. | | | | Answer with a single letter (e.g., A, B, C, ...). **Strictly follow this format–do not include any explanations or additional text.** | | **Answer:** C | **Example for Task 6: Scientific Knowledge Understanding.**
| | |:---| | **Prompt:** | | You are given a passage from a scientific paper that describes part of the policy problem motivating the research. | | | | Summarize the specific issue mentioned in the passage using policy-brief style sentences. Your output should: | | \- Highlight the problem or challenge described | | \- Use accessible, non-technical language (technical terms are allowed when necessary) | | \- Focus only on what is present in the passage | | | | Scientific Text: | | Research shows that restrictive immigration policies and practices are associated with poor health, but far less is known about the relationship between inclusive immigration policies and health. Using data from the United States natality files, we estimate associations between state laws granting undocumented immigrants access to driver’s licenses and perinatal outcomes among 4,047,067 singleton births to Mexican and Central American immigrant birthing people (2008-2021). Fitting multivariable log binomial and linear models, we find that the implementation of a license law is associated with improvements in low birthweight and mean birthweight. Replicating these analyses among U.S.-born non-Hispanic White birthing people, we find no association between the implementation of a license law and birthweight. These findings support the hypothesis that states’ extension of legal rights to immigrants improves the health of the next generation. | | | | Summary: | | **Answer:** None (Use Gemini-2.5-Pro as the judge). | **Example for Task 7: Policy Problem Summarization.**
| | |:---| | **Prompt:** | | You are given a passage from a scientific paper that describes part of the research findings. | | | | Summarize the specific findings using policy-brief style sentences. Your output should: | | \- Clearly state the result(s) presented in the passage | | \- Use plain and direct language (technical terms are allowed when necessary) | | \- Focus only on what is present in the passage | | | | Scientific Text: | | Table 2 displays the results of the ordered logistic regression analyses of community distrust and subjective social isolation. Model 1 shows that, independent of the controls, respondents in 2020 evidenced a significantly increased risk of community distrust. Being a respondent in the 2020 survey was associated with almost 50% greater odds of reporting a higher level of distrust than being a respondent in the 2019 sample. However, these between-wave differences did not vary by age; the interaction between wave of survey and age in Model 2 is not significant. | | | | Summary: | | **Answer:** None (Use Gemini-2.5-Pro as the judge). | **Example for Task 8: Research Findings Summarization.**
| | |:---| | **Prompt:** | | You are given a passage from a scientific paper that describes part of the research study’s methodology. | | | | Summarize the method or approach using policy-brief style sentences. Your output should: | | \- Describe the model, data, or procedure mentioned in the passage | | \- Use clear and accessible language (technical terms are allowed when necessary) | | \- Focus only on what is present in the passage | | | | Scientific Text: | | The survey data were collected by the Energy Institute at the Johannes Kepler University Linz, following high European Union standards of data protection and voluntary study participation. The methodology used in this paper does not require institutional ethical approval according to the guidelines set out by the Energy Institute at the Johannes Kepler University Linz. Confidentiality and anonymity of participants were ensured, and informed written consent was obtained from all the interviewees. | | | | Summary: | | **Answer:** None (Use Gemini-2.5-Pro as the judge). | **Example for Task 9: Study Methods Summarization.**
| | |:---| | **Prompt:** | | You are given a passage from a scientific paper that describes part of the policy implications of the research findings. | | | | Summarize the implications using policy-brief style sentences. Your output should: | | \- Explain how the described result or observation could inform or influence policy | | \- Use accessible language (technical terms are allowed when necessary) | | \- Focus only on what is present in the passage | | | | Scientific Text: | | This study was conducted in Chicago, Illinois. Although these data were collected before recreational marijuana use was legalized in Illinois, it is possible that attitudes toward marijuana use are generally more permissive in Chicago relative to other parts of the country. Future research should examine regional variations in responses to substance use in perinatal care settings. In addition, this analysis is premised on the uptake of legal tasks by individual providers in settings where testing is based on provider discretion. Studies examining how law shapes provider-patient interactions in systems where universal drug testing is carried out may reveal different patterns and outcomes. Another limitation of this study is that the data reflect providers’ interpretations and perceptions of hospital testing protocol. In the future, researchers should examine official, documented hospital protocols to understand how formal organizational policy shapes provider practices. Finally, given sample size limitations and the wide range of state and organizational responses to substance use during pregnancy, these findings are not intended to capture practices in perinatal care by and large. Rather, these findings reflect dilemmas workers may contend with in settings where they are given latitude over the degree to which they adopt legal tasks or not. | | | | Summary: | | **Answer:** None (Use Gemini-2.5-Pro as the judge). | **Example for Task 10: Policy Implications Summarization.**
| | |:---| | **Prompt:** You are given the full text of a scientific research paper. | | | | Identify and summarize the **policy problem** that this research addresses. Your response should: | | \- Clearly state the societal, environmental, or strategic issue motivating the study | | \- Be written in accessible, policy-brief style sentences | | \- Focus on the core challenge the research aims to solve | | \- Avoid excessive detail or technical jargon unless necessary for clarity | | | | Write a short paragraph suitable for inclusion under the “Policy Problem” section of a policy brief. | | | | Scientific Research Paper: | | {**Scientific paper text**} | | | | Policy Problem: | | **Answer:** The rapid diffusion of battery electric vehicles — in addition to the decarbonization of the energy sector — requires an increasing number of batteries. However, the EU’s goal to cover 90% of its battery demand from domestic production by 2030 is at risk, as projected demand will likely exceed 1.0 TWh per year and outpace production capacity despite highly ambitious growth rates. If Europe fails to scale up production, it may face severe economic and geopolitical risks, due to increased dependence on external suppliers, weakened industrial competitiveness and potential for delayed decarbonization. An urgent question is therefore whether Europe can realistically meet its future battery demand through domestic production, and what policy actions are needed to ensure success. | **Example for Task 11: Policy Problem Generation.**
| | |:---| | **Prompt:** You are given the full text of a scientific research paper. | | | | Summarize the paper’s key **research findings** in the style of a policy brief. Your response should: | | \- Focus on the major empirical or model-based results | | \- Present findings that are policy-relevant and supported by the study | | \- Use concise, non-technical language with essential technical terms as needed | | \- Avoid quoting directly or summarizing minor details | | | | Write a short, clear paragraph appropriate for the “Scientific Research Findings” section of a policy brief. | | | | Scientific Research Paper: | | {**Scientific paper text**} | | | | Scientific Research Findings: | | **Answer:** We find that European battery cell demand will likely surpass 1.0 TWh per year by 2030, whereas domestic production capacity is expected to fall short, creating a risk of supply constraints. Although Europe can be expected to meet at least 50-60% of its demand through domestic production by 2030, achieving the EU’s 90% self-sufficiency target is feasible but uncertain, as nearly half of our modelled scenarios fail to meet this target. If Europe wants more independence from battery cell imports, our findings highlight the urgency of accelerating production capacity expansion, scaling up a battery supply chain, and implementing strong industrial policies to support competitiveness and supply sovereignty. Our approach is broadly applicable to regions aiming for battery self-sufficiency and should be examined with interacting factors such as policy support and supply chain resilience. However, our analysis does not account for disruptive market shifts, policy reversals, or unexpected technological breakthroughs, which could substantially alter production and demand trajectories. | **Example for Task 12: Research Findings Generation.**
| | |:---| | **Prompt:** You are given the full text of a scientific research paper. | | | | Summarize the **research methods** used in the study in a way that is informative for policy audiences. Your response should: | | \- Describe the overall study design, data sources, tools, and models used | | \- Avoid unnecessary technical details or equations | | \- Use accessible but precise language | | Write a short paragraph suitable for the “Scientific Research Study Methods” section of a policy brief. | | | | Scientific Research Paper: | | {**Scientific paper text**} | | | | Scientific Research Study Methods: | | **Answer:** Our study uses probabilistic modelling to project future battery demand and domestic production in Europe and evaluates Europe’s pathway towards battery self-sufficiency via probabilistic statements. We independently model demand and supply using S-shaped diffusion curves based on historical data, actual announced production capacities, and practice-oriented findings about how these announced capacities materialize over time. Our approach accounts for uncertainties such as construction delays, utilization rates and evolving market conditions, and assesses corresponding raw material needs. This study is particularly relevant given Europe’s policy push toward climate neutrality and striving for resilient, sustainable battery value chains with domestic production and global competitiveness. By applying an established technology diffusion framework, we provide a robust, scenario-based outlook rather than relying on overly optimistic industry projections. Our method is suitable for evaluating long-term industrial transformation and supply chain resilience, making it applicable to other regions with similar ambitions. | **Example for Task 13: Study Methods Generation.**
Example for Task 14: Policy Implications Generation.
Prompt: You are given the full text of a scientific research paper.
Summarize the study’s policy implications—how the research findings can inform policy or decision-making. Your response should:
- Interpret what the results suggest for governments, regulators, or institutions
- Highlight potential actions, strategies, or decisions informed by the study
- Use policy-appropriate language without speculation beyond the study’s conclusions
- Be clear, practical, and informative
Write a short paragraph for the “Policy Implications” section of a policy brief.
Scientific Research Paper:
{Scientific paper text}
Policy Implications:

Answer:

  • Focus on net materialized production capacities rather than mere corporate announcements to ensure realistic policy planning and avoid overestimations, while ensuring a minimal level of production by local companies utilizing domestic intellectual property.

  • Create predictable and reliable framework conditions for industry and end users to stimulate market demand and allow capacity announcements to materialize.

  • Strengthen public-private partnerships to de-risk investments and streamline European regulations to accelerate the scale-up of battery production and regional supply chains.

  • Create a competitive differentiation and level playing field via de-risking industrial policies, sustainability criteria, and local content requirements at the European level, and carefully balance trade policies to foster competitiveness with options for global collaboration and learning.

  • Ensure continuous policy support and cascade research and development policies with industrial policies in terms of timing and scope in accordance with the scale-up of battery production and an evolving battery value chain.
| | |:---| | **Prompt:** You are given the full text of a scientific research paper. | | | | Your task is to generate a structured **policy brief** that includes the following components: | | \- A concise, descriptive **title** that captures the central theme or policy issue addressed in the paper (do not label it with “Title:”). | | \- **Policy Problem** - What societal or strategic challenge is the paper addressing? | | \- **Scientific Research Findings** - What are the key results relevant to this issue? | | \- **Scientific Research Study Methods** - What methodology or data supports these findings? | | \- **Policy Implications** - How might the findings guide or influence policy? | | | | Write each section in a clear, accessible style suitable for policymakers. Avoid overly technical language, but use essential terms when necessary. Structure the output as a labeled five-part policy brief document. | | | | Scientific Research Paper: | | {**Scientific paper text**} | | | | Policy Brief: | | **Answer:** {**Full Policy brief text**} | **Example for Task 15: Policy Brief Generation.**
| | |:---| | **Prompt:** | | You are given a passage from a scientific paper and a research finding derived from it. | | | | Your task is to determine whether the finding is **fully supported** by the information provided in the passage. | | | | Respond with one word: | | **SUPPORT** - if the finding is clearly and accurately justified by the passage | | **CONTRADICT** - if the finding misrepresents, exaggerates, or is not derivable from the passage | | | | Do not make assumptions beyond the provided text. Use only the given evidence. | | | | Scientific Text: | | Fuelled by substantial BEV diffusion up to 2035, European battery demand is likely to surpass 1.0 TWhyr-1 by 2030 (in 69% of all scenarios). The interquartile range (IQR) in 2030 is 0.97-1.2 TWh yr-1. Some high-demand scenarios may exceed the 1 TWh threshold as early as 2026 and even approach 1.6 TWh yr-1 by 2030, with the top 10% exceeding 1.30 TWh yr-1. | | | | Claimed Finding: | | In around 69% of model runs, European battery demand exceeds 1 TWh per year by 2030. | | | | Answer with SUPPORT or CONTRADICT. | | | | **Strictly follow this format–do not include any explanations or additional text.** | | | | Answer (SUPPORT / CONTRADICT): | | **Answer:** SUPPORT | **Example for Task 16: Scientific Claims Verification.**
| | |:---| | **Prompt:** | | You will be shown a claim related to the COVID-19 pandemic and the abstract of a biomedical research paper. Each sentence from the abstract will be on a separate line. Your task is to return a JSON object with two fields: | | | | \- "verdict": The fact-checking verdict. If the information in the abstract supports the claim, write "SUPPORT". If the abstract contradicts the claim, write "CONTRADICT". | | \- "evidence": An array of sentences providing evidence for the verdict. Please copy all relevant sentences verbatim from the abstract. | | | | For instance, if the model were given the claim "the COVID vaccine is safe for healthy adults", the output might be | | { | |   "verdict": "SUPPORT", | |   "evidence": \["Our findings indicate that the risks of severe COVID vaccine side effects are low for healthy adults."\] | | } | | | | Claim: The d614g mutation in the sars-cov2 spike protein reduces infectivity in an ace2 receptor-dependent manner | | | | Abstract: | | The SARS-CoV2 coronavirus responsible for the current COVID19 pandemic has been reported to have a relatively low mutation rate. | | Nevertheless, a few prevalent variants have arisen that give the appearance of undergoing positive selection as they are becoming increasingly widespread over time. | | Most prominent among these is the D614G amino acid substitution in the SARS-CoV2 Spike protein, which mediates viral entry. | | The D614G substitution, however, is in linkage disequilibrium with the ORF1b P314L mutation, where both mutations almost invariably co-occur, making functional inferences problematic. | | In addition, the possibility of repeated new introductions of the mutant strain does not allow one to distinguish between a founder effect and an intrinsic genetic property of the virus. | | Here, we synthesized and expressed the WT and D614G variant SARS-Cov2 Spike protein, and report that using a SARS-CoV2 Spike protein pseudotyped lentiviral vector we observe that the D614G variant Spike has \>1/2 log(10) increased infectivity in human cells expressing the human ACE2 protein as the viral receptor. | | The increased binding/fusion activity of the D614G Spike protein was corroborated in a cell fusion assay using Spike and ACE2 proteins expressed in different cells. | | These results are consistent with the possibility that the Spike D614G mutant increases the infectivity of SARS-CoV2. | | **Answer:** | | { | |   "verdict": "CONTRADICT", | |   "evidence": \[ | |     "Most prominent among these is the D614G amino acid substitution in the SARS-CoV2 Spike protein, which mediates viral entry.", | |     "The increased binding/fusion activity of the D614G Spike protein was corroborated in a cell fusion assay using Spike and ACE2 proteins expressed in different cells.", | |     "These results are consistent with the possibility that the Spike D614G mutant increases the infectivity of SARS-CoV2." | |   \] | | } | **Example for Task 17: Scientific Claims Verification 2.**
| | |:---| | **Prompt:** You are given the full text of a scientific research paper and a proposed policy implication. | | | | Your task is to determine whether the policy implication is **fully supported** by the scientific research paper. | | | | Respond with one word: | | **SUPPORT** - if the implication clearly follows from the paper’s results and interpretations | | **CONTRADICT** - if the implication misrepresents, overstates, or is not justified by the evidence or conclusions | | | | Base your judgment only on what the paper states. Do not speculate beyond the provided content. | | | | Full Scientific Paper: | | {**Scientific paper text**} | | | | Policy Implication: | | If Europe wants more independence from battery cell imports, our findings highlight the urgency of accelerating production capacity expansion, scaling up a battery supply chain, and implementing strong industrial policies to support competitiveness and supply sovereignty. | | | | Answer with SUPPORT or CONTRADICT. | | | | **Strictly follow this format—do not include any explanations or additional text.** | | | | Answer (SUPPORT / CONTRADICT): | | **Answer:** SUPPORT | **Example for Task 18: Policy Implications Verification.**
### Prompts for Sci2Pol-Bench Evaluation and Dataset Curation We provide detailed prompts used for LLM evaluation and dataset curation for Sci2Pol-Bench. #### Tasks 7-10 Prompt for Reference-free Score We present the detailed prompt used for Gemini-2.5-Pro when serving as the evaluation judge for Tasks 7-10 in 28. We average the scores from the JSON output and multiply the result by 20 to scale it to a 0–100 range.
| | |:---| | **Prompt:** | | You are a strict and critical evaluator of summaries. Evaluate the summary on the following dimensions using a 1-5 scale (1 = very poor, 5 = excellent). Be conservative in your judgments: do not give high scores unless the summary is genuinely outstanding. | | | | \(1\) **Clarity**: whether the summary is reader-friendly and expresses ideas clearly. | | \(2\) **Accuracy**: whether the summary contains the same information as the source document. | | \(3\) **Coverage**: how well the summary covers the important information from the source document. | | \(4\) **Overall quality**: how good the summary is overall at representing the source document; a good summary is a shorter piece of text that has the essence of the original and tries to convey the same information as the source document. | | | | Return only a JSON object in this format: | | { | |   `"clarity"`: \<1-5\>, | |   `"accuracy"`: \<1-5\>, | |   `"coverage"`: \<1-5\>, | |   `"overall_quality"`: \<1-5\> | | } | | | | – | | | | **Source Passage:** | | {**source passage text**} | | | | **Summary:** | | {**summary text**} | **Prompt for LLM-based Judge for Summarization Tasks.**
#### Task 11 Prompt for Reference-based Score We present the detailed prompt used for Gemini-2.5-Pro when serving as the evaluation judge for Tasks 11 (*Policy Problem Generation*) in 29 and 30. Let `prob_imp` and `prob_qual` be their respective JSON outputs, and define the component set ``` math \begin{equation*} C=\{\text{background},\,\text{existing\_problem},\,\text{consequence},\,\text{attention\_problem},\,\text{supporting\_detail}\}. \end{equation*} ``` We compute the raw score of Task 11 as ``` math \begin{equation*} S_{\mathrm{raw}} = \sum_{c\in C} \mathrm{prob\_imp}[c]\cdot \mathrm{prob\_qual}[c] \end{equation*} ``` We then multiply the raw score by 20 to scale it to a maximum value of 100.
| | |:---| | **Prompt:** | | You are a strict policy-brief evaluator. Given the full scientific PAPER, assign an **importance score** to each structural component for effectively communicating the policy problem, based only on the **PAPER**. | | | | **Components:** | | \(1\) **background** — what drives the problem (e.g., scientific, environmental, or economic context). | | \(2\) **existing_problem** — the current obstacle, mismatch, or challenge. | | \(3\) **consequence** — potential risks if the problem is not addressed. | | \(4\) **attention_problem** — the key policy issue or question requiring urgent attention. | | \(5\) **supporting_detail** — clarification or elaboration of any of the above. | | | | **Scoring Instructions:** | | • Assign an importance score between 0.0 and 1.0 for each component. | | • A higher score means the component is essential for understanding the policy problem described in the PAPER. | | • A lower score means the component is optional, minor, or not clearly relevant. | | • If a component is not justified by the PAPER, assign 0.0. | | | | **Strict-grading Instructions:** | | • Score conservatively: if unsure, choose the lower score. | | • Base each score **only** on the PAPER—no external references or assumptions. | | • Return exactly the JSON object below (no explanations, no extra keys). | | | | Return only a JSON object in this format: | | { | |   `"background"`: \<0.0-1.0\>, | |   `"existing_problem"`: \<0.0-1.0\>, | |   `"consequence"`: \<0.0-1.0\>, | |   `"attention_problem"`: \<0.0-1.0\>, | |   `"supporting_detail"`: \<0.0-1.0\> | | } | | | | – | | | | **PAPER:** | | {**full PAPER text**} | **Prompt for LLM-based Judge for Task 11: Policy Problem Generation (Importance).**
| | |:---| | **Prompt:** | | You are a strict policy-brief evaluator. Given the full scientific PAPER and the CANDIDATE’s policy problem paragraph, assign **quality scores** to five aspects of how well the problems are conveyed in **CANDIDATE_POLICY_PROBLEM**. | | | | **Components:** | | \(1\) **background** — what drives the problem (e.g., scientific, environmental, or economic context). | | \(2\) **existing_problem** — the current obstacle, mismatch, or challenge. | | \(3\) **consequence** — potential risks if the problem is not addressed. | | \(4\) **attention_problem** — the key policy issue or question requiring urgent attention. | | \(5\) **supporting_detail** — clarification or elaboration of any of the above. | | | | **Scoring Instructions:** | | • Assign a quality score between 0.0 and 1.0 for each component. | | • A higher score means the content is clear, logical, and strongly aligned with the PAPER. | | • A lower score means the content is vague, incorrect, poorly structured, or missing. | | • If a component is not addressed, assign 0.0. | | | | **Strict-grading Instructions:** | | • Score conservatively: if unsure, choose the lower score. | | • Base each score **only** on comparisons between PAPER and CANDIDATE. | | • Return exactly the JSON object below (no explanations, no extra keys). | | • Only evaluate content in CANDIDATE_POLICY_PROBLEM. | | | | **Return only a JSON object in this format:** | | { | |   `"background"`: \<0.0-1.0\>, | |   `"existing_problem"`: \<0.0-1.0\>, | |   `"consequence"`: \<0.0-1.0\>, | |   `"attention_problem"`: \<0.0-1.0\>, | |   `"supporting_detail"`: \<0.0-1.0\> | | } | | | | – | | | | **PAPER:** | | {**full PAPER text**} | | | | **CANDIDATE_POLICY_PROBLEM:** | | {**candidate policy problem paragraph**} | **Prompt for LLM-based Judge for Task 11: Policy Problem Generation (Quality).**
#### Task 12 Prompt for Reference-based Score We present the detailed prompt used for Gemini-2.5-Pro when serving as the evaluation judge for Task 12 (*Research Findings Generation*) in 31. We average the scores from the JSON output and multiply the result by 100 to scale it to a 0–100 range.
| | |:---| | **Prompt:** | | You are a strict policy-brief evaluator. Given the full scientific PAPER and the CANDIDATE’s findings section, assign **quality scores** to five aspects of how well the findings are conveyed in **CANDIDATE_FINDINGS**. | | | | **Criteria:** | | \(1\) **completeness** — does the section include *all important findings* from the PAPER? | | \(2\) **importance** — are the findings mentioned actually *important* according to the PAPER? | | \(3\) **accuracy** — are the described findings *factually correct* and consistent with the PAPER? | | \(4\) **summarizing_findings** — does the section effectively *emphasize and summarize* the key messages or implications from the data, rather than just listing facts? | | \(5\) **specification_to_findings** — does the section *clarify the scope, context, or limitations* of the findings, including conditions under which they apply? | | | | **Scoring Instructions:** | | • Assign a score between 0.0 and 1.0 for each criterion. | | • A higher score means the section performs well on that criterion. | | • A lower score means the section is vague, misleading, incomplete, or missing that dimension. | | | | **Strict-grading Instructions:** | | • Score conservatively: if unsure, choose the lower score. | | • Base each score **only** on comparisons between PAPER and CANDIDATE. | | • Return exactly the JSON object below (no explanations, no extra keys). | | | | **Return only a JSON object in this format:** | | { | |   `"completeness"`: \<0.0-1.0\>, | |   `"importance"`: \<0.0-1.0\>, | |   `"accuracy"`: \<0.0-1.0\>, | |   `"summarizing_findings"`: \<0.0-1.0\>, | |   `"specification_to_findings"`: \<0.0-1.0\> | | } | | | | – | | | | **PAPER:** | | {**full PAPER text**} | | | | **CANDIDATE_FINDINGS:** | | {**candidate findings section**} | **Prompt for LLM-based Judge for Task 12: Research Findings Generation.**
#### Task 13 Prompt for Reference-based Score We present the detailed prompt used for Gemini-2.5-Pro when serving as the evaluation judge for Task 13 (*Study Methods Generation*) in 32. We use a weighted rubric with the scores from the JSON output. We give greater weight to the first two criteria, because they carry more information, while the third serves only as an auxiliary signal. ``` math \begin{align*} S_{\mathrm{raw}} =& 2 \times \text{\texttt{methods\_clarity\_and\_purpose}} \\ & + 2 \times \text{\texttt{methods\_technicality\_appropriateness}} \\ & + \text{\texttt{methods\_explanation\_of\_terms}} \end{align*} ``` We average the above score by 5 and multiply the result by 100 to scale it to a 0–100 range.
| | |:---| | **Prompt:** | | You are a strict policy-brief evaluator. Given the full scientific PAPER and the CANDIDATE’s methods section, assign **quality scores** to three core aspects of how the methodology is described. Each score should be a float between 0.0 and 1.0. | | | | **Criteria:** | | \(1\) **clarity_and_purpose** — Is the method described in a clear, structured way that highlights *what was done and why*, rather than simply listing tools or data sources? | | \(2\) **technicality_appropriateness** — Is the level of technical detail appropriate for a policy audience without excessive jargon, complexity, or irrelevant detail? | | \(3\) **explanation_of_terms** — Are technical terms, models, or data sources *explained* in accessible language and context without unexplained acronyms or unclear references? | | | | **Scoring Instructions:** | | • Assign a score between 0.0 and 1.0 for each criterion. | | • A higher score means the section performs well on that criterion. | | • A lower score means the section is vague, overly technical, unexplained, or missing that dimension. | | | | **Strict-grading Instructions:** | | • Score conservatively: if unsure, choose the lower score. | | • Base each score **only** on comparisons between PAPER and CANDIDATE. | | • Return exactly the JSON object below (no explanations, no extra keys). | | | | **Return only a JSON object in this format:** | | { | |   `"clarity_and_purpose"`: \<0.0-1.0\>, | |   `"technicality_appropriateness"`: \<0.0-1.0\>, | |   `"explanation_of_terms"`: \<0.0-1.0\> | | } | | | | – | | | | **PAPER:** | | {**full PAPER text**} | | | | **CANDIDATE_METHOD:** | | {**candidate methods section**} | **Prompt for LLM-based Judge for Task 13: Study Methods Generation.**
#### Task 14 Prompt for Reference-based Score We present the detailed prompt used for Gemini-2.5-Pro when serving as the evaluation judge for Task 14 (*Policy Implications Generation*) in 33. We average the scores from the JSON output and multiply the result by 100 to scale it to a 0–100 range.
| | |:---| | **Prompt:** | | You are a **strict** policy-brief evaluator. Given the full scientific PAPER and the CANDIDATE’s policy implications section, assign **quality scores** to the following four criteria. | | | | **Dimensions:** | | \(1\) **accuracy** — Are the implications *explicitly supported* by the PAPER without speculative or hallucinated claims? | | \(2\) **coverage** — Does the section capture *all major implications* stated in the PAPER? | | \(3\) **conciseness_and_distinctness** — Are the implications *concise and non-redundant*? Each point should make a *distinct* contribution. | | \(4\) **alignment_with_paper_intent** — Does the implication reflect the PAPER’s *main message or takeaway* (e.g., recommendation, warning, scientific insight, call to awareness)? | | | | **Scoring Instructions:** | | • Assign a score between 0.0 and 1.0 for each dimension. | | • A higher score means the section performs well on that dimension. | | • A lower score means the section is vague, incorrect, redundant, or misaligned. | | | | **Strict-grading Instructions:** | | • Score conservatively: if unsure, choose the lower score. | | • Base each score **only** on comparisons between PAPER and CANDIDATE. | | • Return exactly the JSON object below (no explanations, no extra keys). | | | | **Return only a JSON object in this format:** | | { | |   `"accuracy"`: \<0.0-1.0\>, | |   `"coverage"`: \<0.0-1.0\>, | |   `"conciseness_and_distinctness"`: \<0.0-1.0\>, | |   `"alignment_with_paper_intent"`: \<0.0-1.0\> | | } | | | | – | | | | **PAPER:** | | {**full PAPER text**} | | | | **CANDIDATE_IMPLICATION:** | | {**candidate policy implications section**} | **Prompt for LLM-based Judge for Task 14: Policy Implications Generation.**
#### Task 15 Prompt for Reference-based Score We present the detailed prompt used for Gemini-2.5-Pro when serving as the evaluation judge for Task 15 (*Policy Brief Generation*) in 34. We average the scores from the JSON output and multiply the result by 100 to scale it to a 0–100 range.
| | |:---| | **Prompt:** | | You are a strict policy-brief evaluator. Given the full scientific PAPER, an EXPERT-written reference brief, and a CANDIDATE brief, grade the CANDIDATE on four dimensions and produce a compact JSON report. | | | | **Evaluation dimensions & conservative 0-5 rubric** | | Start each score at 0 and add points only when the brief clearly meets the criterion. Reserve 4 or 5 for near-flawless performance; 3 means “solid but with notable gaps”; 2 or below signals clear problems. | | 0 = disastrous = poor = fair = good = very good = excellent | | \(1\) **ContextualDepth**: Does the CANDIDATE capture the study’s essential quantitative findings, methods, and broader context (e.g., raw-material outlook, scenario count) *without missing key facts or adding fluff*? | | \(2\) **HallucinationRisk**: Are *all* claims traceable to the PAPER (or universally known)? Deduct heavily for any unsupported number or causal claim. | | \(3\) **ReadabilityTone**: Is the brief concise, logically ordered, written in active voice, and appropriate for policymakers? Penalize lengthy sentences or jargon. | | \(4\) **Actionability**: Are policy implications concrete, tied directly to evidence, and immediately useful? Vague or speculative advice $`\leq`$ 2. | | | | **Output format (MUST be valid JSON; numeric scores only, no prose):** | | { | |   `"contextual_depth"`: \<0-5\>, | |   `"hallucination_risk"`: \<0-5\>, | |   `"readability_tone"`: \<0-5\>, | |   `"actionability"`: \<0-5\> | | } | | | | **Strict-grading instructions:** | | • Score conservatively: if unsure, choose the lower score. | | • Base each score only on comparisons between PAPER and CANDIDATE; EXPERT_BRIEF is reference context. | | • Return exactly the JSON object above (no explanations, no extra keys). | | | | – | | | | **PAPER:** | | {**full PAPER text**} | | | | **EXPERT_BRIEF:** | | {**expert-written brief**} | | | | **CANDIDATE_BRIEF:** | | {**candidate brief**} | **Prompt for LLM-based Judge for Task 15: Policy Brief Generation.**
#### Task 5 Prompt for Data Curation We present the detailed prompt used with GPT-o3 when curating the dataset for Task 5 in 35.
| | |:---| | **Prompt:** | | You are given a scientific paper and a corresponding policy brief. The policy brief includes four components: | | 1\. Policy Problem | | 2\. Scientific Research Findings | | 3\. Scientific Research Study Methods | | 4\. Policy Implications | | | | Your task is to extract valuable, content-rich passages from the scientific paper that correspond to each of these components, **as reflected in the policy brief**. Each sample should preferably contain **three or more coherent and consecutive sentences**, copied **verbatim** from the scientific paper. However, shorter excerpts are acceptable if they are highly informative. Avoid random, trivial, or disjointed selections. | | | | Assign one of the following five labels to each extracted sample: | | \(1\) Policy Problem | | \(2\) Scientific Research Findings | | \(3\) Scientific Research Study Methods (e.g., experimental design, data sources, modeling, and implementation details) | | \(4\) Policy Implications | | \(5\) None: for content unrelated to the policy translation task (including acknowledgments, author contributions, and institutional affiliations) | | | | Return exactly: | | \- 1 sample for Policy Problem | | \- 5 samples for Scientific Research Findings | | \- 5 samples for Scientific Research Study Methods | | \- 2 samples for Policy Implications | | \- 2 samples for None | | | | Output the result as a JSON array of objects, each with the following fields: | | \- `"label"`: one of \["Policy Problem", "Scientific Research Findings", "Scientific Research Study Methods", "Policy Implications", "None"\] | | \- `"text"`: the extracted passage copied verbatim from the scientific paper | | | | **Do not** paraphrase. **Do not** include commentary. Only output a valid JSON array of labeled, **verbatim** text segments. | **Prompt for the Data Curation in Task 5.**
#### Task 11 Prompt for Data Curation We present the detailed prompt used with GPT-o3 when curating the dataset for Task 11 in 36.
| | |:---| | **Prompt:** | | In the following, you will see three examples. Each example includes a scientific research paper and a paragraph describing the **policy problem** that the research addresses, as written for a policy brief. | | | | Your task is to write a new **policy problem** paragraph for a different scientific paper that I will provide. | | | | **Note:** I will also give you an additional paragraph related to the policy problem for the new paper. You may refer to it for context, but it is **not** the desired output. | | | | **Example 1:** | | Scientific paper: {**scientific paper text**} | | Policy problem paragraph: {**policy problem text**} | | | | **Example 2:** | | Scientific paper: {**scientific paper text**} | | Policy problem paragraph: {**policy problem text**} | | | | **Example 3:** | | Scientific paper: {**scientific paper text**} | | Policy problem paragraph: {**policy problem text**} | | | | **New Paper:** {**scientific paper text**} | | Related paragraph for policy problem: {**related policy problem text**} | | Policy problem paragraph: | **Prompt for the Data Curation in Task 11.**
#### Task 13 Prompt for Data Curation We present the detailed prompt used with GPT-o3 when curating the dataset for Task 13 in 37.
| | |:---| | **Prompt:** | | In the following, you will see three examples. Each example includes a scientific research paper and a paragraph describing the **scientific research study method** that the research addresses, as written for a policy brief. | | | | Your task is to write a new **scientific research study method** paragraph for a different scientific paper that I will provide. | | | | **Note:** I will also give you an additional paragraph related to the scientific research study method for the new paper. You may refer to it for context, but it is **not** the desired output. | | | | **Example 1:** | | Scientific paper: {**scientific paper text**} | | Scientific research study method paragraph: {**scientific research study method text**} | | | | **Example 2:** | | Scientific paper: {**scientific paper text**} | | Scientific research study method paragraph: {**scientific research study method text**} | | | | **Example 3:** | | Scientific paper: {**scientific paper text**} | | Scientific research study method paragraph: {**scientific research study method text**} | | | | **New Paper:** {**scientific paper text**} | | Related paragraph for scientific research study method: {**scientific research study method text**} | | Scientific research study method paragraph: | **Prompt for the Data Curation in Task 13.**
#### Task 16 Prompt for Data Curation We present the detailed prompt used with GPT-o3 when curating the dataset for Task 16 in 38.
| | |:---| | **Prompt:** | | You are given a scientific research paper. | | | | Your task is to generate **ten** query-answer pairs for the following binary classification task: | | | | \> Determine whether a stated research finding is *fully supported* by the research results reported in the scientific paper. | | | | Each query-answer pair must include: | | 1\. **research_results** - Copy **one or two consecutive paragraphs verbatim** from the paper that present empirical findings, statistics, or core observations. | | 2\. **research_finding** - Write a concise sentence that either: | | \- *Accurately follows* from the results (**SUPPORT**) | | \- *Sounds plausible*, but is **not actually supported**, misstates causal direction, overgeneralizes, or infers something beyond the evidence (**CONTRADICT**) | | 3\. **answer** - Either `"SUPPORT"` or `"CONTRADICT"` | | | | **Requirements:** | | \- Return **exactly 10 entries** in total. | | \- Include **5** SUPPORT and **5** CONTRADICT examples–no more, no fewer. | | \- Use a *different* results passage for each entry–**do not reuse**. | | \- Make the distinction between SUPPORT and CONTRADICT **subtle and challenging** (e.g., include plausible misinterpretations, causal reversals, or logical overextensions). | | \- Reproduce **all paper text exactly as written**–no paraphrasing, truncation, or ellipses. | | \- Output only a valid **JSON file** containing a list of 10 dictionaries. | | \- Each dictionary must contain exactly the following keys: `"research_results"`, `"research_finding"`, and `"answer"`. | | | | **Output JSON format:** | | {\[ | |   { | |     "research_results": "\", | |     "research_finding": "\", | |     "answer": "SUPPORT" "CONTRADICT" | |   }, | |   ... | | \]} | **Prompt for the Data Curation in Task 16.**
#### Task 18 Prompt for Data Curation We present the detailed prompt used with GPT-o3 when curating the dataset for Task 18 in 39.
| | |:---| | **Prompt:** | | Please rewrite each of the following policy recommendations to express the opposite meaning as clearly and thoroughly as possible. | | | | **Policy Implications:** | | {**policy implication text**} | **Prompt for the Data Curation in Task 18.**
### Prompts for Sci2Pol-Corpus Curation In this section, we present the detailed prompts used for Sci2Pol-Corpus curation: (i) the coarse-grained filtering prompt in 3.2 (40); (ii) the fine-grained filtering prompt in 3.2 (41); and (iii) the in-context polishing prompt in 3.3 (42).
Prompt for the Coarse-grained Filtering Step in 3.2
Prompt:

I will give you a policy document and a scientific article abstract. Your task is to evaluate whether the policy document is primarily about the scientific article it cites. Consider the following criteria:

1. Discussion of the Article’s Content: The policy document must explicitly discuss the findings, methodology, or conclusions of the scientific article in detail.

2. Policy Implications: The document must connect the scientific article to policy decisions, recommendations, or implications for policymakers.

3. Central Focus: The scientific article should be a key focus of the policy document, rather than being just one of many references or a minor supporting citation.

Evaluation Steps:

- Read the scientific article abstract to understand its key points.

- Analyze the policy document to determine whether it engages with the article’s content, its implications, and whether the article is a central focus.

- Score the policy document on the following dimensions:

  • Relevance (0-5): How central is the scientific article to the policy document? (0 = only briefly mentioned, 5 = core focus).

  • Depth of Discussion (0-5): To what extent does the policy document engage with the scientific article’s content (e.g., findings, methodology, conclusions)? (0 = minimal detail, 5 = in-depth discussion).

  • Policy Connection (0-5): How well does the policy document translate the scientific article into policy implications or recommendations? (0 = no connection, 5 = strong, explicit connection).

  • Citation Frequency & Emphasis (0-5): How frequently and prominently is the article referenced in the policy document? (0 = one minor mention, 5 = referenced throughout as a key source).

- Provide a final verdict on whether the policy document is primarily about the scientific article.

- Return the output in a valid JSON format.
Output JSON format:
{[
  {
    "verdict": "Yes" "No",
    "scores": {
      "relevance": 0-5,
      "depth_of_discussion": 0-5,
      "policy_connection": 0-5,
      "citation_frequency_emphasis": 0-5
    },
    "justification": "<3-5 sentence explanation>"
  }
]}
Prompt for the Fine-grained Filtering Step in 3.2
Prompt:

I will give you a policy document and a scientific article. Your task is to evaluate whether the policy document is primarily about the scientific article. Consider the following criteria:

1. Discussion of the Article’s Content: The policy document must explicitly discuss the findings, methodology, or conclusions of the scientific article in detail.

2. Policy Implications: The document must connect the scientific article to policy decisions, recommendations, or implications for policymakers.

3. Central Focus: The scientific article should be a key focus of the policy document, rather than being just one of many references or a minor supporting citation.

Evaluation Steps:

- Read the scientific article to understand its key points.

- Analyze the policy document to determine whether it engages with the article’s content, its implications, and whether the article is a central focus.

- Score the policy document on the following dimensions:

  • Relevance (0-5): How central is the scientific article to the policy document? (0 = only briefly mentioned, 5 = core focus).

  • Depth of Discussion (0-5): To what extent does the policy document engage with the scientific article’s content (e.g., findings, methodology, conclusions)? (0 = minimal detail, 5 = in-depth discussion).

  • Policy Connection (0-5): How well does the policy document translate the scientific article into policy implications or recommendations? (0 = no connection, 5 = strong, explicit connection).

  • Citation Frequency & Emphasis (0-5): How frequently and prominently is the article referenced in the policy document? (0 = one minor mention, 5 = referenced throughout as a key source)

  • Document Similarity (0-5): Are the policy document and the scientific article almost exactly the same with only minor formatting differences? (0 = the text of policy document is very different from the scientific article, 5 = the two documents are nearly identical).

- Provide a final verdict on whether the policy document is primarily about the scientific article.

- Return the output in a valid JSON format.
Output Format:
{[
  {
    "verdict": "Yes" "No",
    "scores": {
      "relevance": 0-5,
      "depth_of_discussion": 0-5,
      "policy_connection": 0-5,
      "citation_frequency_emphasis": 0-5,
      "doc_similarity": 0-5
    },
    "justification": "<3-5 sentence explanation>"
  }
]}
| | |:---| | **Prompt:** | | You are a professional editor specializing in policy briefs based on scientific research. Use the sample scientific papers and their corresponding policy briefs as the standard for tone, structure, and formatting. Based on this reference, revise the draft policy brief for the target scientific paper. Ensure the revised brief is clear, accurate, concise, and policy-relevant, matching the quality of the samples. | | | | Sample Scientific Paper 1:  {**scientific paper text**} | | Sample Policy Brief 1:  {**policy brief text**} | | | | Sample Scientific Paper 2:  {**scientific paper text**} | | Sample Policy Brief 2:  {**policy brief text**} | | | | Sample Scientific Paper 3:  {**scientific paper text**} | | Sample Policy Brief 3:  {**policy brief text**} | | | | Target Scientific Paper:  {**scientific paper text**} | | Draft Policy Brief: {**policy brief text**} | | | | Respond with the **revised policy brief only**, using the following format: | | \- **Policy Problem**: Concise and precise, aligned with sample quality. | | \- **Scientific Research Findings**: Comprehensive and coherent (no bullet points), matching the structure of the original paper. | | \- **Scientific Research Study Methods**: Narrative format (no point form), at the same level of generality and technicality as the samples. | | \- **Policy Implications**: Bullet points only; grounded strictly in the paper’s findings without speculation or external examples. | | | | **Requirements:** | | \- Maintain the same **functional length** as the samples: each section should be long enough to reflect the depth and structure of the specific paper, not artificially extended or shortened to match sample length. Do not pad with filler, overexplain to match longer samples, or oversimplify to match shorter ones. | | \- Use a professional, policy-oriented voice for a scientifically literate audience. | | \- Ensure strict factual alignment with the target scientific paper. | **Prompt for the In-context Polishing Step in 3.3.**
Achiam, Josh, Steven Adler, Sandhini Agarwal, et al. 2023. “Gpt-4 Technical Report.” *arXiv Preprint arXiv:2303.08774*.
Agiza, Ahmed, Mohamed Mostagir, and Sherief Reda. 2024. “PoliTune: Analyzing the Impact of Data Selection and Fine-Tuning on Economic and Political Biases in Large Language Models.” *Proceedings of the 2024 AAAI/ACM Conference on AI, Ethics, and Society*, 2–12.
Anderson, Kathryn Freeman, and Darra Ray-Warren. 2022a. “Racial-Ethnic Residential Clustering and Early COVID-19 Vaccine Allocations in Five Urban Texas Counties.” *Journal of Health and Social Behavior* 63 (4): 472–90. .
Anderson, Kathryn Freeman, and Darra Ray-Warren. 2022b. “Racial-Ethnic Residential Clustering and Early COVID-19 Vaccine Allocations in Five Urban Texas Counties.” *Journal of Health and Social Behavior* 63 (4): 472–90. .
Andreoni, Pietro, Johannes Emmerling, and Massimo Tavoni. 2024a. “Financing Negative Emissions Leads to Windfall Profits and Inequality at Net Zero.” *Nature Climate Change* 14 (1): 20–21.
Andreoni, Pietro, Johannes Emmerling, and Massimo Tavoni. 2024b. “Inequality Repercussions of Financing Negative Emissions.” *Nature Climate Change* 14 (1): 48–54.
Anshassi, Malak, and Timothy G Townsend. 2025a. “Improving Waste Systems in the Global South to Tackle International Environmental Impacts.” *Nature Sustainability*, 1–11.
Anshassi, Malak, and Timothy G Townsend. 2025b. “The Broader Benefits of Waste System Investments in the Global South: Sustainable Materials.” *Nature Sustainability*, 1–2.
Apostoleris, Harry, Sgouris Sgouridis, Marco Stefancich, and Matteo Chiesa. 2018. “Evaluating the Factors That Led to Low-Priced Solar Electricity Projects in the Middle East.” *Nature Energy* 3 (12): 1109–14.
Apostoleris, Harry, Sgouris Sgouridis, Marco Stefancich, and Matteo Chiesa. 2019. “Utility Solar Prices Will Continue to Drop All over the World Even Without Subsidies.” *Nature Energy* 4 (10): 833–34.
Arnautu, Diana, and Christian Dagenais. 2021. “Use and Effectiveness of Policy Briefs as a Knowledge Transfer Tool: A Scoping Review.” *Humanities and Social Sciences Communications* 8 (1): 1–14.
Augustine, Jennifer March. 2021a. “Mothers’ Out-of-Sequence Postsecondary Education and Their Health and Health Behaviors.” *Journal of Health and Social Behavior* 62 (1): 2–18. .
Augustine, Jennifer March. 2021b. “Mothers’ Out-of-Sequence Postsecondary Education and Their Health and Health Behaviors.” *Journal of Health and Social Behavior* 62 (1): 2–18. .
Azar, Ariel. 2024a. “Work–Family Life Course Trajectories and Women’s Mental Health: The Moderating Role of Defamilization Policies in 15 European Territories.” *Journal of Health and Social Behavior* 65 (4): 468–88. .
Azar, Ariel. 2024b. “Work–Family Life Course Trajectories and Women’s Mental Health: The Moderating Role of Defamilization Policies in 15 European Territories.” *Journal of Health and Social Behavior* 65 (4): 468–88. .
Azarova, Valeriya, Dominik Engel, Cornelia Ferner, Andrea Kollmann, and Johannes Reichl. 2018. “Exploring the Impact of Network Tariffs on Household Electricity Expenditures Using Load Profiles and Socio-Economic Characteristics.” *Nature Energy* 3 (4): 317–25.
Azarova, Valeriya, Dominik Engel, Cornelia Ferner, Andrea Kollmann, and Johannes Reichl. 2019. “Transition to Peak-Load-Based Tariffs Can Be Disruptive for Different Groups of Consumers.” *Nature Energy* 4 (10): 829–30.
Bai, Hui, Jan G Voelkel, Shane Muldowney, Johannes C Eichstaedt, and Robb Willer. 2025. “LLM-Generated Messages Can Persuade Humans on Policy Issues.” *Nature Communications* 16 (1): 6037.
Bai, Jinze, Shuai Bai, Yunfei Chu, et al. 2023. “Qwen Technical Report.” *arXiv Preprint arXiv:2309.16609*.
Barrie, Christopher, Alexis Palmer, and Arthur Spirling. 2024. “Replication for Language Models Problems, Principles, and Best Practice for Political Science.” *URL: Https://Arthurspirling. Org/Documents/BarriePalmerSpirling TrustMeBro. Pdf*.
Basheer, Mohammed, Victor Nechifor, Alvaro Calzadilla, et al. 2023a. “Cooperative Adaptive Management of the Nile River with Climate and Socio-Economic Uncertainties.” *Nature Climate Change* 13 (1): 48–57.
Basheer, Mohammed, Victor Nechifor, Alvaro Calzadilla, et al. 2023b. “Negotiating Nile Infrastructure Management Should Consider Climate Change Uncertainties.” *Nature Climate Change* 13 (1): 17–19.
Bechtel, Michael M, Kenneth F Scheve, and Elisabeth van Lieshout. 2020. “Constant Carbon Pricing Increases Support for Climate Action Compared to Ramping up Costs over Time.” *Nature Climate Change* 10 (11): 1004–9.
Bechtel, Michael M, Kenneth F Scheve, and Elisabeth van Lieshout. 2021. “Most People Prefer Constant Carbon Costs over Increasing Cost Schedules Even If Costs Are High.” *Nature Climate Change* 11 (11): 909–10.
Berg, Mark T, Ethan M Rogers, Man-Kit Lei, and Ronald L Simons. 2021a. “Losing Years Doing Time: Incarceration Exposure and Accelerated Biological Aging Among African American Adults.” *Journal of Health and Social Behavior* 62 (4): 460–76. .
Berg, Mark T, Ethan M Rogers, Man-Kit Lei, and Ronald L Simons. 2021b. “Losing Years Doing Time: Incarceration Exposure and Accelerated Biological Aging Among African American Adults.” *Journal of Health and Social Behavior* 62 (4): 460–76. .
Bierman, Alex, and Scott Schieman. 2020a. “Social Estrangement and Psychological Distress Before and During the COVID-19 Pandemic: Patterns of Change in Canadian Workers.” *Journal of Health and Social Behavior* 61 (4): 398–417. .
Bierman, Alex, and Scott Schieman. 2020b. “Social Estrangement and Psychological Distress Before and During the COVID-19 Pandemic: Patterns of Change in Canadian Workers.” *Journal of Health and Social Behavior* 61 (4): 398–417. .
Bjørn, Anders, Shannon M Lloyd, Matthew Brander, and H Damon Matthews. 2022a. “Renewable Energy Certificates Allow Companies to Overstate Their Emission Reductions.” *Nature Climate Change* 12 (6): 508–9.
Bjørn, Anders, Shannon M Lloyd, Matthew Brander, and H Damon Matthews. 2022b. “Renewable Energy Certificates Threaten the Integrity of Corporate Science-Based Targets.” *Nature Climate Change* 12 (6): 539–46.
Bommasani, Rishi, Sanjeev Arora, Jennifer Chayes, et al. 2025. “Advancing Science-and Evidence-Based AI Policy.” *Science* 389 (6759): 459–61.
Bonan, Jacopo, Cristina Cattaneo, Giovanna d’Adda, and Massimo Tavoni. 2020a. “Combining Information on Others’ Energy Usage and Their Approval of Energy Conservation Promotes Energy Saving Behaviour.” *Nature Energy* 5 (11): 832–33.
Bonan, Jacopo, Cristina Cattaneo, Giovanna d’Adda, and Massimo Tavoni. 2020b. “The Interaction of Descriptive and Injunctive Social Norms in Promoting Energy Conservation.” *Nature Energy* 5 (11): 900–909.
Braunholtz-Speight, Tim, Maria Sharmina, Edward Manderson, et al. 2020a. “Business Models and Financial Characteristics of Community Energy in the UK.” *Nature Energy* 5 (2): 169–77.
Braunholtz-Speight, Tim, Maria Sharmina, Edward Manderson, et al. 2020b. “Price Support Allows Communities to Raise Low-Cost Citizen Finance for Renewable Energy Projects.” *Nature Energy* 5 (2): 127–28.
Brynjolfsson, E, A Korinek, and A Agrawal. 2024. *The Economics of Transformative AI: A Research Agenda.”*. Stanford Digital Economy Lab. Draft manuscript.
Buck, Holly Jean, Wim Carton, Jens Friis Lund, and Nils Markusson. 2023a. “Countries’ Long-Term Climate Strategies Fail to Define Residual Emissions.” *Nature Climate Change* 13 (4): 317–19.
Buck, Holly Jean, Wim Carton, Jens Friis Lund, and Nils Markusson. 2023b. “Why Residual Emissions Matter Right Now.” *Nature Climate Change* 13 (4): 351–58.
Budolfson, Mark, Francis Dennig, Frank Errickson, et al. 2021a. “Climate Action with Revenue Recycling Has Benefits for Poverty, Inequality and Well-Being.” *Nature Climate Change* 11 (12): 1111–16.
Budolfson, Mark, Francis Dennig, Frank Errickson, et al. 2021b. “Protecting the Poor with a Carbon Tax and Equal Per Capita Dividend.” *Nature Climate Change* 11 (12): 1025–26.
Caggiano, Holly, Sara M Constantino, Chris Greig, and Elke U Weber. 2024a. “Community Benefits Can Build Bipartisan Support for Large-Scale Energy Infrastructure.” *Nature Energy* 9 (10): 1187–88.
Caggiano, Holly, Sara M Constantino, Chris Greig, and Elke U Weber. 2024b. “Public and Local Policymaker Preferences for Large-Scale Energy Project Characteristics.” *Nature Energy* 9 (10): 1230–40.
Cai, Hao, Jiaqi Xuan, Xiaoxi Wang, et al. 2025. “The Multiple Benefits of Chinese Dietary Transformation.” *Nature Sustainability*, 1–13.
Calzada, Igor, Géza Németh, and Mohammed Salah Al-Radhi. 2025. “Trustworthy AI for Whom? GenAI Detection Techniques of Trust Through Decentralized Web3 Ecosystems.” *Big Data and Cognitive Computing* 9 (3): 62.
Casey, Joan A, Jason G Su, Lucas RF Henneman, et al. 2020a. “Coal-Fired Power Plant Closures and Retrofits Reduce Asthma Morbidity in the Local Population.” *Nature Energy* 5 (5): 365–66.
Casey, Joan A, Jason G Su, Lucas RF Henneman, et al. 2020b. “Improved Asthma Outcomes Observed in the Vicinity of Coal Power Plant Retirement, Retrofit and Conversion to Natural Gas.” *Nature Energy* 5 (5): 398–408.
Cerf, Moran, Sandra C Matz, and Malcolm A MacIver. 2023a. “Participating in a Climate Futures Market Increases Support for Costly Climate Policies.” *Nature Climate Change* 13 (6): 511–12.
Cerf, Moran, Sandra C Matz, and Malcolm A MacIver. 2023b. “Participating in a Climate Prediction Market Increases Concern about Global Warming.” *Nature Climate Change* 13 (6): 523–31.
Ci, Zhijia, Wenjie Shen, Baowei Chen, et al. 2024a. “Mercury Risk in Blue Carbon Ecosystems.” *Nature Sustainability* 7 (12): 1560–61.
Ci, Zhijia, Wenjie Shen, Baowei Chen, et al. 2024b. “Potential Increase of Neurotoxic Mercury Risk in Global Blue Carbon Nature-Based Solutions.” *Nature Sustainability* 7 (12): 1592–95.
Coz, Pierre Le, Jia An Liu, Debarun Bhattacharjya, Georgina Curto, and Serge Stinckwich. 2025. “What Would an LLM Do? Evaluating Policymaking Capabilities of Large Language Models.” *arXiv Preprint arXiv:2509.03827*.
Czarnecki, Danielle, Danielle Bessett, Hillary J Gyuras, Alison H Norris, and Michelle L McGowan. 2023a. “State of Confusion: Ohio’s Restrictive Abortion Landscape and the Production of Uncertainty in Reproductive Health Care.” *Journal of Health and Social Behavior* 64 (4): 470–85. .
Czarnecki, Danielle, Danielle Bessett, Hillary J Gyuras, Alison H Norris, and Michelle L McGowan. 2023b. “State of Confusion: Ohio’s Restrictive Abortion Landscape and the Production of Uncertainty in Reproductive Health Care.” *Journal of Health and Social Behavior* 64 (4): 470–85. .
Dao, Tri. 2024. “FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning.” *International Conference on Learning Representations*.
Deshmukh, Ranjit, Paige Weber, Olivier Deschenes, et al. 2023a. “Equitable Low-Carbon Transition Pathways for California’s Oil Extraction.” *Nature Energy* 8 (6): 597–609.
Deshmukh, Ranjit, Paige Weber, Olivier Deschenes, et al. 2023b. “Well Setbacks Limit California’s Oil Supply with Larger Health Benefits and Employment Losses Than Excise and Carbon Taxes.” *Nature Energy* 8 (6): 562–64.
Diezmartı́nez, Claudia V, Benjamin K Sovacool, and Anne G Short Gianotti. 2024a. “Implementing Climate Justice in Boston’s Building Performance Standard.” *Nature Cities* 1 (10): 628–30.
Diezmartı́nez, Claudia V, Benjamin K Sovacool, and Anne G Short Gianotti. 2024b. “Operationalizing Climate Justice in the Implementation of Boston’s Building Performance Standard.” *Nature Cities* 1 (10): 665–76.
Dipoppa, Gemma, and Saad Gulzar. 2024. “Bureaucrat Incentives Reduce Crop Burning and Child Mortality in South Asia.” *Nature* 634 (8036): 1125–31.
Dipoppa, Gemma, and Saad Gulzar. 2025. *Air Pollution Could Be Reduced by Incentivizing Local Government Officials to Control Crop Burning*. NATURE PORTFOLIO HEIDELBERGER PLATZ 3, BERLIN, 14197, GERMANY.
Dore, Emily C, Surbhi Shrivastava, and Patricia Homan. 2024a. “Structural Sexism and Preventive Health Care Use in the United States.” *Journal of Health and Social Behavior* 65 (1): 2–19. .
Dore, Emily C, Surbhi Shrivastava, and Patricia Homan. 2024b. “Structural Sexism and Preventive Health Care Use in the United States.” *Journal of Health and Social Behavior* 65 (1): 2–19. .
Druckenmiller, Hannah, Yanjun Liao, Sophie Pesek, Margaret Walls, and Shan Zhang. 2024a. “Removing Development Incentives in Risky Areas Promotes Climate Adaptation.” *Nature Climate Change* 14 (9): 936–42.
Druckenmiller, Hannah, Yanjun Liao, Sophie Pesek, Margaret Walls, and Shan Zhang. 2024b. “Removing Development Incentives in Risky Areas Reduces Climate Damages and Yields Co-Benefits.” *Nature Climate Change* 14 (9): 901–2.
Druckman, James N. 2015. “Communicating Policy-Relevant Science.” *PS: Political Science & Politics* 48 (S1): 58–69.
Duan, Yuwan, Zengkai Zhang, Yuze Li, Shouyang Wang, Cuihong Yang, and Yi Lu. 2024a. “Global Corporate Tax Competition Challenges Climate Change Mitigation.” *Nature Climate Change* 14 (4): 353–56.
Duan, Yuwan, Zengkai Zhang, Yuze Li, Shouyang Wang, Cuihong Yang, and Yi Lu. 2024b. “Global Corporate Tax Competition Leads to Unintended yet Non-Negligible Climate Impacts.” *Nature Climate Change* 14 (4): 314–15.
Egli, Florian, Bjarne Steffen, and Tobias S Schmidt. 2018. “A Dynamic Analysis of Financing Conditions for Renewable Energy Technologies.” *Nature Energy* 3 (12): 1084–92.
Egli, Florian, Bjarne Steffen, and Tobias S Schmidt. 2019. “Learning in the Financial Sector Is Essential for Reducing Renewable Energy Costs.” *Nature Energy* 4 (10): 835–36.
Fang, Zhichao, Jonathan Dudek, Ed Noyons, and Rodrigo Costas. 2024. “Science Cited in Policy Documents: Evidence from the Overton Database.” *arXiv Preprint arXiv:2407.09854*.
Fosch-Villaronga, Eduard, Mohammed Raiz Shaffique, Marie Schwed-Shenker, Antoni Mut-Piña, Simone van der Hof, and Bart Custers. 2025. “Science for Robot Policy: Advancing Robotics Policy Through the EU Science for Policy Approach.” *Technological Forecasting and Social Change* 218: 124202.
Furnas, Alexander C., Timothy M. LaPira, and Dashun Wang. 2025. “Partisan Disparities in the Use of Science in Policy.” *Science* 388 (6745): 362–67. .
Gars, Johan, Daniel Spiro, and Henrik Wachtmeister. 2022a. “European Fuel Tax Cuts Increase Russian Oil Profits.” *Nature Energy* 7 (10): 912–13.
Gars, Johan, Daniel Spiro, and Henrik Wachtmeister. 2022b. “The Effect of European Fuel-Tax Cuts on the Oil Income of Russia.” *Nature Energy* 7 (10): 989–97.
Gasparini, Matteo, Matthew C Ives, Ben Carr, Sophie Fry, and Eric Beinhocker. 2024. “Model-Based Financial Regulations Impair the Transition to Net-Zero Carbon Emissions.” *Nature Climate Change* 14 (5): 476–81.
Gasparini, Matteo, Matthew Ives, Ben Carr, Sophie Fry, and Eric Beinhocker. 2024. “Model-Based Financial Regulation Challenges for the Net-Zero Transition.” *Nature Climate Change* 14 (5): 434–35.
Goldstein, Anna, Claudia Doblinger, Erin Baker, and Laura Dı́az Anadón. 2020a. “Patenting and Business Outcomes for Cleantech Startups Funded by the Advanced Research Projects Agency-Energy.” *Nature Energy* 5 (10): 803–10.
Goldstein, Anna, Claudia Doblinger, Erin Baker, and Laura Dı́az Anadón. 2020b. “Startups Supported by ARPA-e Were More Innovative Than Others but an Investment Gap May Remain.” *Nature Energy* 5 (10): 741–42.
Gruber, Katharina, Tobias Gauster, Gregor Laaha, Peter Regner, and Johannes Schmidt. 2022a. “Profitability and Investment Risk of Texan Power System Winterization.” *Nature Energy* 7 (5): 409–16.
Gruber, Katharina, Tobias Gauster, Gregor Laaha, Peter Regner, and Johannes Schmidt. 2022b. “Winterizing Power Plants Pays Off for Risk-Neutral Investors in Texas.” *Nature Energy* 7 (5): 398–99.
Hall, Stephen, Jillian Anable, Jeffrey Hardy, Mark Workman, Christoph Mazur, and Yvonne Matthews. 2021a. “Innovative Energy Business Models Appeal to Specific Consumer Groups but May Exacerbate Existing Inequalities for the Disengaged.” *Nature Energy* 6 (4): 337–38.
Hall, Stephen, Jillian Anable, Jeffrey Hardy, Mark Workman, Christoph Mazur, and Yvonne Matthews. 2021b. “Matching Consumer Segments to Innovative Utility Business Models.” *Nature Energy* 6 (4): 349–61.
Han, Xiaowen, Tom VanHeuvelen, Jeylan T Mortimer, and Zachary Parolin. 2024a. “Cumulative Unionization and Physical Health Disparities Among Older Adults.” *Journal of Health and Social Behavior* 65 (2): 162–81. .
Han, Xiaowen, Tom VanHeuvelen, Jeylan T Mortimer, and Zachary Parolin. 2024b. “Cumulative Unionization and Physical Health Disparities Among Older Adults.” *Journal of Health and Social Behavior* 65 (2): 162–81. .
Harring, Niklas, Erik Jönsson, Simon Matti, Gabriela Mundaca, and Sverker C Jagers. 2023a. “Cross-National Analysis of Attitudes Towards Fossil Fuel Subsidy Removal.” *Nature Climate Change* 13 (3): 244–49.
Harring, Niklas, Erik Jönsson, Simon Matti, Gabriela Mundaca, and Sverker C Jagers. 2023b. “Public Acceptance of Fossil Fuel Subsidy Removal Can Be Reinforced with Revenue Recycling.” *Nature Climate Change* 13 (3): 214–15.
Heerden, Rik van, Oreane Y Edelenbosch, Vassilis Daioglou, et al. 2025a. “Demand-Side Policies Can Significantly Reduce Emissions from Energy Use in Buildings and Transport.” *Nature Energy*, 1–2.
Heerden, Rik van, Oreane Y Edelenbosch, Vassilis Daioglou, et al. 2025b. “Demand-Side Strategies Enable Rapid and Deep Cuts in Buildings and Transport Emissions to 2050.” *Nature Energy* 10 (3): 380–94.
Hermansen, Are Skeie, Andrew Penner, István Boza, et al. 2025a. “Immigrant–Native Pay Gap Driven by Lack of Access to High-Paying Jobs.” *Nature*, 1–7.
Hermansen, Are Skeie, Andrew Penner, István Boza, et al. 2025b. “Pay Gap Between Nationals and Migrants Mainly Due to Unequal Access to High-Paying Jobs.” *Nature*.
Hsu, Pin-Lun, Yun Dai, Vignesh Kothapalli, et al. 2025. “Liger-Kernel: Efficient Triton Kernels for LLM Training.” *Championing Open-Source DEvelopment in ML Workshop @ ICML25*. .
Huang, Yue, Chujie Gao, Siyuan Wu, et al. 2025. “On the Trustworthiness of Generative Foundation Models: Guideline, Assessment, and Perspective.” *arXiv Preprint arXiv:2502.14296*.
Huang, Yue, Zhengqing Yuan, Yujun Zhou, et al. 2024. “Social Science Meets LLMs: How Reliable Are Large Language Models in Social Simulations?” *arXiv Preprint arXiv:2410.23426*.
Ivanov, Valeriy Y, Vinh Ngoc Tran, Weichen Huang, et al. 2024. “Urban Flooding Is Intensified by Outdated Design Guidelines and a Lack of a Systems Approach.” *Nature Cities* 1 (10): 626–27.
Iyer, Gokul, Yang Ou, James Edmonds, et al. 2022a. “Ratcheting of Climate Pledges Needed to Limit Peak Global Warming.” *Nature Climate Change* 12 (12): 1129–35.
Iyer, Gokul, Yang Ou, James Edmonds, et al. 2022b. “The Path to 1.5 c Requires Ratcheting of Climate Pledges.” *Nature Climate Change* 12 (12): 1092–93.
Janssens, Charlotte, Petr Havlı́k, Tamás Krisztin, et al. 2020. “Global Hunger and Climate Change Adaptation Through International Trade.” *Nature Climate Change* 10 (9): 829–35.
Janssens, Charlotte, Petr Havlı́k, Tamás Krisztin, et al. 2021. “International Trade Is a Key Component of Climate Change Adaptation.” *Nature Climate Change* 11 (11): 915–16.
Kar, Abhishek, Shonali Pachauri, Rob Bailis, and Hisham Zerriffi. 2019. “Using Sales Data to Assess Cooking Gas Adoption and the Impact of India’s Ujjwala Programme in Rural Karnataka.” *Nature Energy* 4 (9): 806–14.
Kar, Abhishek, Shonali Pachauri, Rob Bailis, and Hisham Zerriffi. 2020. “Capital Cost Subsidies Through India’s Ujjwala Cooking Gas Programme Promote Rapid Adoption of Liquefied Petroleum Gas but Not Regular Use.” *Nature Energy* 5 (2): 125–26.
Kaufmann, Robert K, and Caitlin Connelly. 2020a. “Non-Market Forces Significantly Affect Oil Prices.” *Nature Energy* 5 (2): 129–30.
Kaufmann, Robert K, and Caitlin Connelly. 2020b. “Oil Price Regimes and Their Role in Price Diversions from Market Fundamentals.” *Nature Energy* 5 (2): 141–49.
Kennedy, Kathleen M, Morgan R Edwards, Claudia Doblinger, et al. 2024a. “Rapid Rise in Corporate Climate-Tech Investments Complements Support from Public Grants.” *Nature Energy* 9 (7): 773–74.
Kennedy, Kathleen M, Morgan R Edwards, Claudia Doblinger, et al. 2024b. “The Effects of Corporate Investment and Public Grants on Climate and Energy Startup Outcomes.” *Nature Energy* 9 (7): 883–93.
Kitzing, Lena, Morten Kofoed Jensen, Thomas Telsnig, and Eric Lantz. 2020a. “Multifaceted Drivers for Onshore Wind Energy Repowering and Their Implications for Energy Transition.” *Nature Energy* 5 (12): 1012–21.
Kitzing, Lena, Morten Kofoed Jensen, Thomas Telsnig, and Eric Lantz. 2020b. “Multifaceted Political and Social Drivers Inform Wind Energy Repowering Decisions and Potential.” *Nature Energy* 5 (12): 950–51.
Kontokosta, Constantine E, Danielle Spiegel-Feld, and Sokratis Papadopoulos. 2020a. “Mandatory Building Energy Audits Alone Are Insufficient to Meet Climate Goals.” *Nature Energy* 5 (4): 282–83.
Kontokosta, Constantine E, Danielle Spiegel-Feld, and Sokratis Papadopoulos. 2020b. “The Impact of Mandatory Energy Audits on Building Energy Use.” *Nature Energy* 5 (4): 309–16.
Lamb, William F, Thomas Gasser, Rosa M Roman-Cuesta, et al. 2024a. “Current National Proposals Are Off Track to Meet Carbon Dioxide Removal Needs.” *Nature Climate Change* 14 (6): 555–56.
Lamb, William F, Thomas Gasser, Rosa M Roman-Cuesta, et al. 2024b. “The Carbon Dioxide Removal Gap.” *Nature Climate Change* 14 (6): 644–51.
Li, Haitao, You Chen, Qingyao Ai, Yueyue Wu, Ruizhe Zhang, and Yiqun Liu. 2024. “Lexeval: A Comprehensive Chinese Legal Benchmark for Evaluating Large Language Models.” *The Thirty-Eighth Annual Conference on Neural Information Processing Systems*.
Li, Lincan, Jiaqi Li, Catherine Chen, et al. 2024. “Political-Llm: Large Language Models in Political Science.” *arXiv Preprint arXiv:2412.06864*.
Li, Sihang, Jin Huang, Jiaxi Zhuang, et al. 2024. “Scilitllm: How to Adapt Llms for Scientific Literature Understanding.” *arXiv Preprint arXiv:2408.15545*.
Lin, Zihang, Yian Yin, Lu Liu, and Dashun Wang. 2023. “SciSciNet: A Large-Scale Open Data Lake for the Science of Science Research.” *Scientific Data* 10 (1): 315.
Link, Steffen, Lara Schneider, Annegret Stephan, Lukas Weymann, and Patrick Plötz. 2025a. “Feasibility of Meeting Future Battery Demand via Domestic Cell Production in Europe.” *Nature Energy*, 1–9.
Link, Steffen, Lara Schneider, Annegret Stephan, Lukas Weymann, and Patrick Plötz. 2025b. “Reliable Industrial Policies Required to Support the Ramp-up of European Battery Production.” *Nature Energy*, 1–2.
Link, Steffen, Annegret Stephan, Daniel Speth, and Patrick Plötz. 2024a. “Declining Costs Imply Fast Market Uptake of Zero-Emission Trucks.” *Nature Energy* 9 (8): 924–25.
Link, Steffen, Annegret Stephan, Daniel Speth, and Patrick Plötz. 2024b. “Rapidly Declining Costs of Truck Batteries and Fuel Cells Enable Large-Scale Road Freight Electrification.” *Nature Energy* 9 (8): 1032–39.
Linsenmeier, Manuel, Adil Mohommad, and Gregor Schwerhoff. 2023a. “Global Benefits of the International Diffusion of Carbon Pricing Policies.” *Nature Climate Change* 13 (7): 679–84.
Linsenmeier, Manuel, Adil Mohommad, and Gregor Schwerhoff. 2023b. “Leadership in Carbon Pricing Encourages Other Countries to Follow.” *Nature Climate Change* 13 (7): 613–14.
Liu, Aixin, Bei Feng, Bing Xue, et al. 2024. “Deepseek-V3 Technical Report.” *arXiv Preprint arXiv:2412.19437*.
Liu, Bo, and Deepak Rajagopal. 2019. “Life-Cycle Energy and Climate Benefits of Energy Recovery from Wastes and Biomass Residues in the United States.” *Nature Energy* 4 (8): 700–708.
Liu, Jiateng, Zhenhailong Wang, Xiaojiang Huang, et al. 2025. “Analyzing and Internalizing Complex Policy Documents for LLM Agents.” *arXiv Preprint arXiv:2510.11588*.
Liu, Yixin, Alexander Richard Fabbri, Jiawen Chen, et al. 2024. “Benchmarking Generation and Evaluation Capabilities of Large Language Models for Instruction Controllable Summarization.” *Findings of the Association for Computational Linguistics: NAACL 2024*, 4481–501.
Liu, Yu, Mingxi Du, Lingyu Yang, et al. 2025a. “Mitigation Policies Interactions Delay the Achievement of Carbon Neutrality in China.” *Nature Climate Change* 15 (2): 147–52.
Liu, Yu, Mingxi Du, Lingyu Yang, et al. 2025b. “Policy Interactions Make Achieving Carbon Neutrality in China More Challenging.” *Nature Climate Change* 15 (2): 134–35.
Longden, Thomas, Simon Quilty, Brad Riley, Lee V White, Michael Klerck, Vanessa Napaltjari Davis, et al. 2022. “Energy Insecurity During Temperature Extremes in Remote Australia.” *Nature Energy* 7 (1): 43–54.
Longden, Thomas, Simon Quilty, Brad Riley, Lee V White, Michael Klerck, Vanessa Napaltjarri Davis, et al. 2022. “Temperature Extremes Exacerbate Energy Insecurity for Indigenous Communities in Remote Australia.” *Nature Energy* 7 (1): 11–12.
Lu, Yifan, K Supriya, Shanna Shaked, Elizabeth H Simmons, and Alexander Kusenko. 2025. “Incentivizing Supplemental Math Assignments and Using AI-Generated Hints Is Associated with Improved Exam Performance.” *Physical Review Physics Education Research* 21 (1): 010160.
Mahdavi, Paasha, Michael L Ross, and Evelyn Simoni. 2025a. “Fossil Fuel Subsidy Reforms Have Become More Fragile.” *Nature Climate Change*, 1–6.
Mahdavi, Paasha, Michael L Ross, and Evelyn Simoni. 2025b. “Government Efforts to Reduce Fossil Fuel Subsidies Have Failed at a Very High Rate.” *Nature Climate Change*, 1–2.
Mani, Sunil, Abhishek Jain, Saurabh Tripathi, and Carlos F Gould. 2020a. “Sustained LPG Use Requires Progress on Broader Development Outcomes.” *Nature Energy* 5 (6): 430–31.
Mani, Sunil, Abhishek Jain, Saurabh Tripathi, and Carlos F Gould. 2020b. “The Drivers of Sustained Use of Liquified Petroleum Gas in India.” *Nature Energy* 5 (6): 450–57.
Manzer, Jamie L, and Ann V Bell. 2021a. “‘We’re a Little Biased’: Medicine and the Management of Bias Through the Case of Contraception.” *Journal of Health and Social Behavior* 62 (2): 120–35. .
Manzer, Jamie L, and Ann V Bell. 2021b. “‘We’re a Little Biased’: Medicine and the Management of Bias Through the Case of Contraception.” *Journal of Health and Social Behavior* 62 (2): 120–35. .
Masters, Ryan K, Andrea M Tilstra, Daniel H Simon, and Kate Coleman-Minahan. 2023a. “Differences in Determinants: Racialized Obstetric Care and Increases in US State Labor Induction Rates.” *Journal of Health and Social Behavior* 64 (2): 174–91. .
Masters, Ryan K, Andrea M Tilstra, Daniel H Simon, and Kate Coleman-Minahan. 2023b. “Differences in Determinants: Racialized Obstetric Care and Increases in US State Labor Induction Rates.” *Journal of Health and Social Behavior* 64 (2): 174–91. .
Mays, Jacob, David P Morton, and Richard P O’Neill. 2019a. “Asymmetric Risk and Fuel Neutrality in Electricity Capacity Markets.” *Nature Energy* 4 (11): 948–56.
Mays, Jacob, David P Morton, and Richard P O’Neill. 2019b. “Decarbonizing Electricity Requires Re-Evaluating Capacity Mechanisms.” *Nature Energy* 4 (11): 912–13.
McCabe, Katharine. 2022a. “Criminalization of Care: Drug Testing Pregnant Patients.” *Journal of Health and Social Behavior* 63 (2): 162–76. .
McCabe, Katharine. 2022b. “Criminalization of Care: Drug Testing Pregnant Patients.” *Journal of Health and Social Behavior* 63 (2): 162–76. .
McCarty, Nolan, Keith T Poole, and Howard Rosenthal. 2016. *Polarized America: The Dance of Ideology and Unequal Riches*. Mit Press.
McFarland, Michael J, Terrence D Hill, and Jennifer Karas Montez. 2023a. “Income Inequality and Population Health: Examining the Role of Social Policy.” *Journal of Health and Social Behavior* 64 (1): 2–20. .
McFarland, Michael J, Terrence D Hill, and Jennifer Karas Montez. 2023b. “Income Inequality and Population Health: Examining the Role of Social Policy.” *Journal of Health and Social Behavior* 64 (1): 2–20. .
Mehta, Piyush, Marc Muller, Meredith T Niles, and Kyle Frankel Davis. 2025. “Child Diet Diversity and Irrigation Expansion in the Global South.” *Nature Sustainability*, 1–9.
Mehta, Piyush, Marc Müller, Meredith T Niles, and Kyle Frankel Davis. 2025. “Nutritional Outcomes of Irrigation Expansion: Agriculture and Nutrition.” *Nature Sustainability*, 1–2.
Merfort, Leon, Nico Bauer, Florian Humpenöder, et al. 2023a. “Bioenergy-Induced Land-Use-Change Emissions with Sectorally Fragmented Policies.” *Nature Climate Change* 13 (7): 685–92.
Merfort, Leon, Nico Bauer, Florian Humpenöder, et al. 2023b. “State of Global Land Regulation Inadequate to Control Biofuel Land-Use-Change Emissions.” *Nature Climate Change* 13 (7): 610–12.
Mildenberger, Matto, Erick Lachapelle, Kathryn Harrison, and Isabelle Stadelmann-Steffen. 2022a. “Limited Evidence That Carbon Tax Rebates Have Increased Public Support for Carbon Pricing.” *Nature Climate Change* 12 (2): 121–22.
Mildenberger, Matto, Erick Lachapelle, Kathryn Harrison, and Isabelle Stadelmann-Steffen. 2022b. “Limited Impacts of Carbon Tax Rebate Programmes on Public Support for Carbon Pricing.” *Nature Climate Change* 12 (2): 141–47.
Millinger, Markus, Fredrik Hedenus, E Zeyen, Fabian Neumann, L Reichenberg, and Göran Berndes. 2025. “Diversity of Biomass Usage Pathways to Achieve Emissions Targets in the European Energy System.” *Nature Energy* 10 (2): 226–42.
Millinger, Markus, F Hedenus, E Zeyen, F Neumann, L Reichenberg, and G Berndes. 2025. “Biomass Exclusion Must Be Weighed Against Benefits of Carbon Supply in European Energy System.” *Nature Energy* 10 (2): 159–61.
Minten, Hannah, Julian Hausweiler, Benedict Probst, Christiane Reinert, Raoul Meys, and André Bardow. 2025a. “Closing Emission Gaps in Border Carbon Adjustments for Chemicals and Plastics: Climate Policy and Trade.” *Nature Sustainability*, 1–2.
Minten, Hannah, Julian Hausweiler, Benedict Probst, Christiane Reinert, Raoul Meys, and André Bardow. 2025b. “Embodied Emissions of Chemicals Within the EU Carbon Border Adjustment Mechanism.” *Nature Sustainability*, 1–10.
Mintrom, Michael, Philippa Goddard, and Shanti Sumartojo. 2025. “Anticipatory Policy Design for Robots in Public Spaces.” *Journal of Asian Public Policy*, 1–15.
Moffette, Fanny, Jennifer Alix-Garcia, Katherine Shea, and Amy H Pickens. 2021a. “Freely Available Deforestation Alerts Can Reduce Emissions from Land-Use Change.” *Nature Climate Change* 11 (11): 913–14.
Moffette, Fanny, Jennifer Alix-Garcia, Katherine Shea, and Amy H Pickens. 2021b. “The Impact of Near-Real-Time Deforestation Alerts Across the Tropics.” *Nature Climate Change* 11 (2): 172–78.
Moinester, Margot, and Kaitlyn K Stanhope. 2024a. “Extending Driver’s Licenses to Undocumented Immigrants: Comparing Perinatal Outcomes Following This Policy Shift.” *Journal of Health and Social Behavior* 65 (3): 324–39. .
Moinester, Margot, and Kaitlyn K Stanhope. 2024b. “Extending Driver’s Licenses to Undocumented Immigrants: Comparing Perinatal Outcomes Following This Policy Shift.” *Journal of Health and Social Behavior* 65 (3): 324–39. .
Mollborn, Stefanie, Jennifer A Pace, and Bethany Rigles. 2025a. “Children’s Health Lifestyles and the Perpetuation of Inequalities.” *Journal of Health and Social Behavior* 66 (1): 2–17. .
Mollborn, Stefanie, Jennifer A Pace, and Bethany Rigles. 2025b. “Children’s Health Lifestyles and the Perpetuation of Inequalities.” *Journal of Health and Social Behavior* 66 (1): 2–17. .
Monteiro, Joao, Pierre-Andre Noel, Etienne Marcotte, et al. 2024. “RepLiQA: A Question-Answering Dataset for Benchmarking LLMs on Unseen Reference Content.” *The Thirty-Eighth Annual Conference on Neural Information Processing Systems*.
Motoki, Fabio, Valdemar Pinho Neto, and Victor Rodrigues. 2024. “More Human Than Human: Measuring ChatGPT Political Bias.” *Public Choice* 198 (1): 3–23.
Mou, Xinyi, Jingcong Liang, Jiayu Lin, et al. 2025. “AgentSense: Benchmarking Social Intelligence of Language Agents Through Interactive Scenarios.” *Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)*, 4975–5001.
Navia Simon, Daniel, and Laura Diaz Anadon. 2025a. “Faster Deployment of Renewables Stabilizes Electricity Prices in Europe.” *Nature Energy*, 1–2.
Navia Simon, Daniel, and Laura Diaz Anadon. 2025b. “Power Price Stability and the Insurance Value of Renewable Technologies.” *Nature Energy* 10 (3): 329–41.
Nguyen, Huyen, Haihua Chen, Lavanya Pobbathi, and Junhua Ding. 2024. “A Comparative Study of Quality Evaluation Methods for Text Summarization.” *arXiv Preprint arXiv:2407.00747*.
Nowak, Andreea C, Lucy Njuguna, Julian Ramirez-Villegas, Pytrik Reidsma, Krystal Crumpler, and Todd S Rosenstock. 2024a. “Enhanced Policy Adequacy Facilitates National Climate Adaptation Tracking Across Africa.” *Nature Climate Change* 14 (8): 787–88.
Nowak, Andreea C, Lucy Njuguna, Julian Ramirez-Villegas, Pytrik Reidsma, Krystal Crumpler, and Todd S Rosenstock. 2024b. “Opportunities to Strengthen Africa’s Efforts to Track National-Level Climate Adaptation.” *Nature Climate Change* 14 (8): 876–82.
O’Shaughnessy, Eric, Galen Barbose, Sudha Kannan, and Jenny Sumner. 2024a. “Community Solar Reaches Adopters Underserved by Rooftop Solar.” *Nature Energy* 9 (8): 926–27.
O’Shaughnessy, Eric, Galen Barbose, Sudha Kannan, and Jenny Sumner. 2024b. “Evaluating Community Solar as a Measure to Promote Equitable Clean Energy Access.” *Nature Energy* 9 (8): 955–63.
Odenweller, Adrian, and Falko Ueckerdt. 2025a. “An Adjusted Strategy Is Needed to Ground Green Hydrogen Expectations in Reality.” *Nature Energy* 10 (1): 19–20.
Odenweller, Adrian, and Falko Ueckerdt. 2025b. “The Green Hydrogen Ambition and Implementation Gap.” *Nature Energy* 10 (1): 110–23.
Ogier, Emily M, Gretta T Pecl, Terry Hughes, et al. 2025a. “Enhance Responsible Governance to Match the Scale and Pace of Marine–Climate Interventions.” *Nature Climate Change*, 1–2.
Ogier, Emily M, Gretta T Pecl, Terry Hughes, et al. 2025b. “Novel Marine-Climate Interventions Hampered by Low Consensus and Governance Preparedness.” *Nature Climate Change*, 1–10.
Owens, Jayanti. 2020a. “Social Class, Diagnoses of Attention-Deficit/Hyperactivity Disorder, and Child Well-Being.” *Journal of Health and Social Behavior* 61 (2): 134–52. .
Owens, Jayanti. 2020b. “Social Class, Diagnoses of Attention-Deficit/Hyperactivity Disorder, and Child Well-Being.” *Journal of Health and Social Behavior* 61 (2): 134–52. .
Pachauri, Shonali, Miguel Poblete-Cazenave, Arda Aktas, and Matthew J Gidden. 2021a. “Access to Clean Cooking Services in Energy and Emission Scenarios After COVID-19.” *Nature Energy* 6 (11): 1067–76.
Pachauri, Shonali, Miguel Poblete-Cazenave, Arda Aktas, and Matthew J Gidden. 2021b. “Clean Cooking Access May Stall Under Slow Post-Pandemic Recovery and Ambitious Climate Mitigation Without Explicit Focus.” *Nature Energy* 6 (11): 1009–10.
Parbst, Matthew, and Blair Wheaton. 2023a. “The Effect of Welfare State Policy Spending on the Equalization of Socioeconomic Status Disparities in Mental Health.” *Journal of Health and Social Behavior* 64 (3): 336–53. .
Parbst, Matthew, and Blair Wheaton. 2023b. “The Effect of Welfare State Policy Spending on the Equalization of Socioeconomic Status Disparities in Mental Health.” *Journal of Health and Social Behavior* 64 (3): 336–53. .
Peng, Wei, Gokul Iyer, Matthew Binsted, et al. 2021a. “The Surprisingly Inexpensive Cost of State-Driven Emission Control Strategies.” *Nature Climate Change* 11 (9): 738–45.
Peng, Wei, Gokul Iyer, Matthew Binsted, et al. 2021b. “To Achieve Deep Cuts in US Emissions, State-Driven Policy Is Only Slightly More Expensive Than Nationally Uniform Policy.” *Nature Climate Change* 11 (11): 911–12.
Phan, Long, Alice Gatti, Ziwen Han, et al. 2025. “Humanity’s Last Exam.” *arXiv Preprint arXiv:2501.14249*.
Rajagopal, Deepak, and Bo Liu. 2020. “The United States Can Generate up to 3.2 EJ of Energy Annually from Waste.” *Nature Energy* 5 (1): 18–19.
Rapp, Kristen Schorpp, Vanessa V Volpe, Tabitha L Hale, and Dominique F Quartararo. 2022a. “State–Level Sexism and Gender Disparities in Health Care Access and Quality in the United States.” *Journal of Health and Social Behavior* 63 (1): 2–18. .
Rapp, Kristen Schorpp, Vanessa V Volpe, Tabitha L Hale, and Dominique F Quartararo. 2022b. “State–Level Sexism and Gender Disparities in Health Care Access and Quality in the United States.” *Journal of Health and Social Behavior* 63 (1): 2–18. https://doi.org/[10.1177/00221465211073836 ](https://doi.org/10.1177/00221465211073836 ).
Ren, Yuanyi, Haoran Ye, Hanjun Fang, Xin Zhang, and Guojie Song. 2024. “ValueBench: Towards Comprehensively Evaluating Value Orientations and Understanding of Large Language Models.” In *In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)*, edited by Lun-Wei Ku, Andre Martins, and Vivek Srikumar. Association for Computational Linguistics. .
Reuel, Anka, and Trond Arne Undheim. 2024. “Generative AI Needs Adaptive Governance.” *arXiv Preprint arXiv:2406.04554*.
Rinscheid, Adrian, and Rolf Wüstenhagen. 2019a. “German Voters Would Prefer a More Ambitious Timeline to Phase Out Coal.” *Nature Energy* 4 (12): 1016–17.
Rinscheid, Adrian, and Rolf Wüstenhagen. 2019b. “Germany’s Decision to Phase Out Coal by 2038 Lags Behind Citizens’ Timing Preferences.” *Nature Energy* 4 (10): 856–63.
Röttger, Paul, Valentin Hofmann, Valentina Pyatkin, et al. 2024. “Political Compass or Spinning Arrow? Towards More Meaningful Evaluations for Values and Opinions in Large Language Models.” *Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)*, 15295–311.
Ruig, Lars T de, Toon Haer, Hans de Moel, et al. 2022a. “Climate-Proofing the National Flood Insurance Program.” *Nature Climate Change* 12 (11): 975–76.
Ruig, Lars T de, Toon Haer, Hans de Moel, et al. 2022b. “How the USA Can Benefit from Risk-Based Premiums Combined with Flood Protection.” *Nature Climate Change* 12 (11): 995–98.
Ryve, Andreas, and Kirsti Hemmi. 2019. “Educational Policy to Improve Mathematics Instruction at Scale: Conceptualizing Contextual Factors.” *Educational Studies in Mathematics* 102 (3): 379–94.
Schnittker, Jason, and Duy Do. 2020a. “Pharmaceutical Side Effects and Mental Health Paradoxes Among Racial-Ethnic Minorities.” *Journal of Health and Social Behavior* 61 (1): 4–23. .
Schnittker, Jason, and Duy Do. 2020b. “Pharmaceutical Side Effects and Mental Health Paradoxes Among Racial-Ethnic Minorities.” *Journal of Health and Social Behavior* 61 (1): 4–23. .
Shears, Esther, Jonas Meckling, and Jared J Finnegan. 2025a. “How Central Banks Address Climate and Transition Risks.” *Nature Energy*, 1–2.
Shears, Esther, Jonas Meckling, and Jared J Finnegan. 2025b. “How Central Banks Manage Climate and Energy Transition Risks.” *Nature Energy*, 1–9.
Shen, Yelong, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, et al. 2022. “Lora: Low-Rank Adaptation of Large Language Models.” *International Conference on Learning Representations*.
Singh, Amanpreet, Mike D’Arcy, Arman Cohan, Doug Downey, and Sergey Feldman. 2023. “SciRepEval: A Multi-Format Benchmark for Scientific Document Representations.” *Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing*, 5548–66.
Singhal, Karan, Tao Tu, Juraj Gottweis, et al. 2025. “Toward Expert-Level Medical Question Answering with Large Language Models.” *Nature Medicine*, 1–8.
Sitarz, Joanna, Michael Pahle, Sebastian Osorio, Gunnar Luderer, and Robert Pietzcker. 2024a. “EU Carbon Prices Signal High Policy Credibility and Farsighted Actors.” *Nature Energy* 9 (6): 691–702.
Sitarz, Joanna, Michael Pahle, Sebastian Osorio, Gunnar Luderer, and Robert Pietzcker. 2024b. “Policy Credibility Is a Key Component for an Effective and Efficient EU Emissions Trading System.” *Nature Energy* 9 (6): 637–38.
Snow, C. P. 1959. *The Two Cultures and the Scientific Revolution*. Cambridge University Press.
Straf, Miron L, Thomas A Schwandt, and Kenneth Prewitt. 2012. *Using Science as Evidence in Public Policy*. National Academies Press.
Sun, Tao, Yuanhao Feng, Chad Zanocco, June Flora, Arun Majumdar, and Ram Rajagopal. 2025a. “Solar and Batteries Are Affordable Options for US Households: Energy Resilience.” *Nature Energy*, 1–2.
Sun, Tao, Yuanhao Feng, Chad Zanocco, June Flora, Arun Majumdar, and Ram Rajagopal. 2025b. “Solar and Battery Can Reduce Energy Costs and Provide Affordable Outage Backup for US Households.” *Nature Energy*, 1–16.
Szomszor, Martin, and Euan Adie. 2022. “Overton: A Bibliometric Database of Policy Document Citations.” *Quantitative Science Studies* 3 (3): 624–50.
Tang, Ling, Junai Yang, Jiali Zheng, et al. 2024a. “Assessing the Impacts of Fertility and Retirement Policies on China’s Carbon Emissions.” *Nature Climate Change* 14 (12): 1261–67.
Tang, Ling, Junai Yang, Jiali Zheng, et al. 2024b. “Relaxing Fertility Policies and Delaying Retirement Age Increase China’s Carbon Emissions.” *Nature Climate Change* 14 (12): 1228–29.
Taylor, Ross, Marcin Kardas, Guillem Cucurull, et al. 2022. “Galactica: A Large Language Model for Science.” *arXiv Preprint arXiv:2211.09085*.
Team, Gemma, Thomas Mesnard, Cassidy Hardin, et al. 2024. “Gemma: Open Models Based on Gemini Research and Technology.” *arXiv Preprint arXiv:2403.08295*.
Thombs, Ryan P, Dennis L Thombs, Andrew K Jorgenson, and Taylor Harris Braswell. 2020a. “What Is Driving the Drug Overdose Epidemic in the United States?” *Journal of Health and Social Behavior* 61 (3): 275–89. .
Thombs, Ryan P, Dennis L Thombs, Andrew K Jorgenson, and Taylor Harris Braswell. 2020b. “What Is Driving the Drug Overdose Epidemic in the United States?” *Journal of Health and Social Behavior* 61 (3): 275–89. .
Tiefenbeck, Verena, Anselma Wörner, Samuel Schöb, Elgar Fleisch, and Thorsten Staake. 2019a. “Real-Time Feedback Promotes Energy Conservation in the Absence of Volunteer Selection Bias and Monetary Incentives.” *Nature Energy* 4 (1): 35–41.
Tiefenbeck, Verena, Anselma Wörner, Samuel Schöb, Elgar Fleisch, and Thorsten Staake. 2019b. “Real-Time Feedback Reduces Energy Consumption Among the Broader Public Without Financial Incentives.” *Nature Energy* 4 (10): 831–32.
Touvron, Hugo, Thibaut Lavril, Gautier Izacard, et al. 2023. “Llama: Open and Efficient Foundation Language Models.” *arXiv Preprint arXiv:2302.13971*.
Tran, Vinh Ngoc, Valeriy Y Ivanov, Weichen Huang, et al. 2024. “Connectivity in Urbanscapes Can Cause Unintended Flood Impacts from Stormwater Systems.” *Nature Cities* 1 (10): 654–64.
Vendetti, V, LD Comencini, F Deriu, V Modugno, et al. 2025. “Passing the Turing Test in Political Discourse: Fine-Tuning LLMs to Mimic Polarized Social Media Comments.” *arXiv Preprint arXiv:2506.14645*.
Vivier, Lucas, Alessio Mastrucci, and Bas van Ruijven. 2025a. “European Union Needs Large Heat Pump and Targeted Renovation Subsidies to Meet Heating Targets: Climate Policy.” *Nature Climate Change*, 1–2.
Vivier, Lucas, Alessio Mastrucci, and Bas van Ruijven. 2025b. “Meeting Climate Target with Realistic Demand-Side Policies in the Residential Sector.” *Nature Climate Change*, 1–8.
Vuolo, Mike, Laura C Frizzell, and Brian C Kelly. 2022a. “Surveillance, Self-Governance, and Mortality: The Impact of Prescription Drug Monitoring Programs on US Overdose Mortality, 2000–2016.” *Journal of Health and Social Behavior* 63 (3): 337–56. .
Vuolo, Mike, Laura C Frizzell, and Brian C Kelly. 2022b. “Surveillance, Self-Governance, and Mortality: The Impact of Prescription Drug Monitoring Programs on US Overdose Mortality, 2000–2016.” *Journal of Health and Social Behavior* 63 (3): 337–56. .
Wadden, David, Kejian Shi, Jacob Morrison, et al. 2024. “SciRIFF: A Resource to Enhance Language Model Instruction-Following over Scientific Literature.” *The Thirty-Eighth Conference on Neural Information Processing Systems Workshop on Foundation Models for Science: Progress, Opportunities, and Challenges*.
Wang, Dashun, and Albert-László Barabási. 2021. *The Science of Science*. Cambridge University Press.
Wang, Xiaoxi, Hao Cai, Jiaqi Xuan, et al. 2025. “Multi-Benefit Diet Changes in China: Food System Policy.” *Nature Sustainability*, 1–2.
Wang, Yafei, Yuxuan Ye, Robert J Nicholls, Lennart Olsson, Detlef P van Vuuren, et al. 2025. “Development Policy Affects Coastal Flood Exposure in China More Than Sea-Level Rise.” *Nature Climate Change*, 1–7.
Wang, Yafei, Yuxuan Ye, Robert J Nicholls, Lennart Olsson, Detlef P. van Vuuren, et al. 2025. “Managing Development Choices Is Essential to Reduce Coastal Flood Risk in China: Coastal Adaptation.” *Nature Climate Change*, 1–2.
Wang, Yubo, Xueguang Ma, Ge Zhang, et al. 2024. “Mmlu-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark.” *The Thirty-Eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track*.
Weber, Eartha, Vassilus Daioglou, Laszlo Vreedenburgh, et al. 2025a. “Climate Action and Clean Cooking Are Vital for Sustaining Air Pollution Related Health Benefits.” *Nature Sustainability*, 1–2.
Weber, Eartha, Vassilus Daioglou, Laszlo Vreedenburgh, et al. 2025b. “Modelling PM2.5 Reduction Scenarios for Future Cardiopulmonary Disease Reduction.” *Nature Sustainability*, 1–16.
White, Lee V, Bradley Riley, Sally Wilson, et al. 2024a. “Geographies of Regulatory Disparity Underlying Australia’s Energy Transition.” *Nature Energy* 9 (1): 92–105.
White, Lee V, Bradley Riley, Sally Wilson, et al. 2024b. “Regulatory Disparities Disadvantage Remote Australian Communities in Energy Transition.” *Nature Energy* 9 (1): 14–15.
White, Lee V, and Nicole D Sintov. 2020a. “Health and Financial Impacts of Demand-Side Response Measures Differ Across Sociodemographic Groups.” *Nature Energy* 5 (1): 50–60.
White, Lee V, and Nicole D Sintov. 2020b. “Varied Health and Financial Impacts of Time-of-Use Energy Rates Across Sociodemographic Groups Raise Equity Concerns.” *Nature Energy* 5 (1): 16–17.
Wolske, Kimberly S, Annika Todd-Blick, and Emma Tome. 2023a. “Behaviourally-Informed Peer Referral Programmes Can Increase the Reach of Low-Income Energy Policies.” *Nature Energy* 8 (8): 787–88.
Wolske, Kimberly S, Annika Todd-Blick, and Emma Tome. 2023b. “Increasing the Reach of Low-Income Energy Programmes Through Behaviourally Informed Peer Referral.” *Nature Energy* 8 (8): 850–58.
Xie, Chengxing, Canyu Chen, Feiran Jia, et al. 2024. “Can Large Language Model Agents Simulate Human Trust Behavior?” *The Thirty-Eighth Annual Conference on Neural Information Processing Systems*.
Ye, Xiaofeng, Yueming Lucy Qiu, Destenie Nock, and Bo Xing. 2025a. “Heat Pumps Can Help Alleviate Residential Energy Insecurity in the USA: Energy Equity.” *Nature Energy*, 1–2.
Ye, Xiaofeng, Yueming Lucy Qiu, Destenie Nock, and Bo Xing. 2025b. “The Comfort Rebound from Heat Pumps and Impact on Household Cooling Behaviour and Energy Security.” *Nature Energy*, 1–12.
Yin, Yian, Yuxiao Dong, Kuansan Wang, Dashun Wang, and Benjamin F Jones. 2022. “Public Use and Public Funding of Science.” *Nature Human Behaviour* 6 (10): 1344–50.
Younas, Zahid Irshad, Atiqa Qureshi, and Mamdouh Abdulaziz Saleh Al-Faryan. 2022. “Financial Inclusion, the Shadow Economy and Economic Growth in Developing Economies.” *Structural Change and Economic Dynamics* 62: 613–21.
Zeng, Yi, Yu Yang Yang, Andy Zhou, et al. 2024. “AIR-Bench 2024: A Safety Benchmark Based on Risk Categories from Regulations and Policies.” *AGI-Artificial General Intelligence-Robotics-Safety & Alignment* 1 (1).
Zhang, Dan, Ziniu Hu, Sining Zhoubian, et al. 2024. “SciInstruct: A Self-Reflective Instruction Annotated Dataset for Training Scientific Language Models.” *The Thirty-Eighth Annual Conference on Neural Information Processing Systems*.
Zhang, Tianyi, Faisal Ladhak, Esin Durmus, Percy Liang, Kathleen McKeown, and Tatsunori B Hashimoto. 2024. “Benchmarking Large Language Models for News Summarization.” *Transactions of the Association for Computational Linguistics* 12: 39–57.
Zhang, Yiwen, Rongbin Xu, Wenzhong Huang, et al. 2025a. “Health Risks of Exposure to Wildfire-Toxic Air.” *Nature Sustainability* 8 (5): 472–73.
Zhang, Yiwen, Rongbin Xu, Wenzhong Huang, et al. 2025b. “Respiratory Risks from Wildfire-Specific PM2. 5 Across Multiple Countries and Territories.” *Nature Sustainability* 8 (5): 474–84.
Zheng, Yaowei, Richong Zhang, Junhao Zhang, et al. 2024. “LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models.” *Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations)* (Bangkok, Thailand).