Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Toward Scientific Reasoning in LLMs: Training from Expert Discussions via Reinforcement Learning

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that expert CRISPR forum discussions, converted into multiple-choice questions and used as reinforcement-learning training data, teach open-source language models to surpass commercial competitors on a genomics reasoning…

desk verdict A promising pipeline for turning expert forum discussions into RL training data, but the unvalidated LLM-generated benchmark leaves the headline accuracy gains without a solid foundation. read the letter →

arxiv 2505.19501 v2 pith:SDAS6GF2 submitted 2025-05-26 cs.AI

Ming Yin , Yuanhao Qu , Ling Yang , Le Cong , Mengdi Wang This is my paper
classification cs.AI
keywords scientificreasoningreinforcementlearningGRPOGenome-BenchCRISPRexpertdiscussionmultiple-choicegenerationmixture-of-agents
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that a fully automated pipeline can convert a decade of real expert discussions on CRISPR genome engineering into a reinforcement-learning training set, and that doing so measurably improves how open-source language models reason about scientific questions. To support this, the authors build Genome-Bench, 3,332 multiple-choice questions extracted from 11 years of public forum threads, and fine-tune open models with Group Relative Policy Optimization, rewarding both the correct choice and a structured explanation. They report accuracy gains of more than fifteen percentage points over the base models, and a learned router that selects among four fine-tuned experts reaches 81.07% on the benchmark test set, ahead of the best commercial system they compare against. The significance is that messy, contextual expert dialogue—not just curated textbooks or exams—can serve as a scalable training signal for domain-specific scientific reasoning.

What carries the argument

The central machinery is the combination of a conversation-to-question pipeline and rule-based reinforcement learning. The pipeline parses raw .mbox email threads from a CRISPR forum, uses GPT-4-Turbo to extract (question, answer, context) triplets, and uses GPT-4o to rewrite the questions, generate plausible wrong-answer options, and package the correct answer with an expert explanation inside <explanation> and <answer> tags. Training is then driven by Group Relative Policy Optimization (GRPO), a policy-gradient method that samples several candidate outputs for each question and compares their rewards within the group instead of training a value network. The reward is +2 for selecting the correct option, +1 for producing the required explanation-and-answer format, and 0 otherwise; the router is trained separately with +1 when the expert it selects answers correctly and -1 when it does not. This reward design is what lets raw forum data become a training signal for reasoning rather than just a corpus to memorize.

What would settle it

An independent panel of CRISPR researchers, blind to the forum-derived labels, re-answers a random sample of about 100 Genome-Bench test questions; if panel agreement with the benchmark labels is low, or if RL models' accuracy advantage over base models disappears when only panel-verified labels are scored, the central claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that authentic expert discussion is a usable training signal for scientific reasoning: a decade of CRISPR forum threads can be mined into a multiple-choice benchmark, and reinforcement learning on that benchmark teaches open-source language models to answer genomic-reasoning questions at a level competitive with, and in one configuration above, commercial frontier systems. The authors build Genome-Bench (3,332 questions from 11 years of forum posts), fine-tune five open models with GRPO under a rule-based reward that credits both the correct letter and a structured explanation, and train a small router to choose among four fine-tuned experts. They report that every RL-tuned model gains more than fifteen accuracy points over its base version, and the RL-trained router reaches 81.07% on the test set, ahead of DeepSeek-R1's 78.97%; on a 35-question human-evaluation subset the router scores 67.6% versus 73.3% for human experts. They also report that correctness and reasoning quality decouple: questions answered correctly after RL can still be accompanied by partially or entirely wrong explanations.

Load-bearing premise

The whole pipeline rests on the assumption that the answers posted in the CRISPR forum are scientifically correct and that the language models used to extract, rewrite, and create the wrong-answer options do not distort the intended meaning; the paper does not independently verify the labels against expert knowledge.

Editorial extensions

If this is right

  • Fine-tuning with GRPO on Genome-Bench pushes every tested open model up by more than fifteen percentage points over its base instruction-tuned version.
  • An RL-trained router that picks among four fine-tuned experts reaches 81.07% on the test split, beating every individual RL model and the strongest commercial baseline tested (78.97%).
  • RL generalizes better than supervised fine-tuning on this benchmark: SFT overfits the smallest model, while RL keeps train and test performance close.
  • Answer accuracy and reasoning quality decouple after RL: some correct answers come with partially or completely wrong explanations, so outcome-only rewards leave reasoning under-supervised.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the extraction pipeline transfers, any field with an archived expert Q&A corpus—chemistry help lists, clinical consult archives, software issue trackers—could be turned into an RL reasoning benchmark; the bottleneck would be verifying extracted ground-truth answers, not the reinforcement-learning step.
  • The router result points to an alternative to monolithic model scaling: a handful of medium-sized specialists coordinated by a learned router can beat a frontier commercial model on a single domain; whether that holds outside this benchmark is untested.
  • The 1.5B Pass@K crossover suggests RL at small scale may be sharpening the output distribution rather than adding new knowledge; an out-of-domain biology question set would separate those two explanations.
  • A direct extension the authors do not test is a reward term that penalizes explanations contradicting the chosen answer; that would indicate whether reasoning quality, not just answer accuracy, can be optimized.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. This paper presents Genome-Bench, a multiple-choice benchmark of 3,332 questions constructed from CRISPR forum discussions via a GPT-4-based extraction pipeline, and uses GRPO reinforcement learning with a rule-based reward to fine-tune several open-weight LLMs. The authors report accuracy gains of over 15% relative to base models, and an RL-trained router that selects among four expert models reaches 81.07% on Genome-Bench, exceeding commercial baselines including DeepSeek-R1. A human evaluation on 35 questions reports 73.3% average expert accuracy, with the RL-Router at 67.6%. The central claim is that RL on expert discussion data improves scientific reasoning in genomics.

Significance. If the benchmark labels are reliable, this is a valuable contribution: it provides a new domain-specific QA benchmark, a modular data pipeline, and evidence that RL with rule-based rewards can improve MCQ accuracy in a scientific domain. Strengths include the public code, the breadth of models evaluated, and the attempt at qualitative reasoning analysis. However, the significance is conditional on the validity of the automatically generated ground truth, which is not independently established. The small human evaluation and the paper's own qualitative findings of flawed reasoning limit the strength of the 'scientific reasoning' claim.

major comments (4)
  1. [Section 3 (Steps 1-4) and Table 2] The ground-truth labels in Genome-Bench are not independently validated. Section 3 Step 1 uses GPT-4-Turbo to extract (question, answer, context) from forum threads, Step 2 uses GPT-4o to rewrite questions and generate distractors, and Step 4 describes filtering as involving 'human inspection' without reporting who performed it, how many items were checked, or any agreement statistic. The human evaluation in Section 5 uses only 35 questions and yields 73.3% average expert agreement with the gold labels, which is consistent with substantial label noise and does not validate the benchmark. Since the training and test splits come from the same pipeline, the >15% gains in Table 2 could be inflated by learning the synthetic labeling scheme rather than scientific reasoning. Please add an independent expert audit of a random sample of items with inter-annotator agreement, or evaluate on an existing expert-curated benchmark.
  2. [References [45] and Introduction] Reference [45] is the same arXiv identifier as this paper (arXiv:2505.19501). The text states that 'as an companion effort detailed in [45]' Genome-Bench is introduced, but this is a self-citation to the current manuscript, so the purported companion paper does not exist. The benchmark construction must be fully described in this paper or cited to a genuinely separate companion publication; as written, the provenance claim is misleading.
  3. [Section 4.1, 4.2, and Figure 5] The reward in Section 4.1 gives +2 for a correct answer determined by the automatically generated labels and +1 for format, and the same pipeline produces train and test data. The claim that RL improves 'scientific reasoning' is undercut by the paper's own qualitative results: Section 4.2 reports cases of correct answers with flawed reasoning, and Figure 5 shows examples of 'partially wrong' and 'completely wrong' reasoning despite correct answers. Please provide a systematic comparison of reasoning quality, such as human ratings of explanations before and after RL, or rephrase the claimed contribution to improved answer accuracy on Genome-Bench.
  4. [Section 5 and Table 2] The human evaluation is too small and lacks statistical support for the 'narrowing the gap' claim. With 35 questions, the standard error of the 73.3% human accuracy is roughly 7.7 percentage points, so the 5.7-point gap between RL-Router (67.6%) and human (73.3%) is not significant. Similarly, the 81.07% vs 78.97% difference between the RL-Router and DeepSeek-R1 on 661 test items is within about 1.5 standard errors. Please report per-expert scores, confidence intervals, and significance tests for the key comparisons.
minor comments (6)
  1. [Abstract] 'Our code are available' should read 'Our code is available'.
  2. [Main text, Figure 1] The main text contains a duplicate of the Figure 1 caption and an inserted duplicate block after the caption; remove the duplicate.
  3. [Section 3 and Appendix D.1] Section 3 says 'seven thematic categories' but lists eight categories, and Appendix D.1 also lists eight; correct the count or merge two categories consistently.
  4. [Appendix G and Section 5] Appendix G says the 35 questions were 'randomly selected' while Section 5 says they were 'selected' as challenging and diverse; clarify the sampling procedure.
  5. [Section 3, Step 4] The phrase 'verified, high-quality QA items' is not supported by details; specify the verification process and any exclusion statistics.
  6. [Table 2] Consider adding confidence intervals to the main accuracy comparisons, especially for the RL-Router vs DeepSeek-R1 difference.

Circularity Check

1 steps flagged · score 2.0 of 10

One minor self-citation to the paper's own arXiv ID for the benchmark; no central derivation reduces to its inputs.

  1. other [Section 1 (Introduction), first paragraph; Reference [45]]
    "As an companion effort detailed in [ 45], we introduce Genome-Bench, a scientific reasoning benchmark in genomics comprising of questions and answers extracted from over a decade of archived scientific forum discussions among genome editing practitioners [ 45]."

    Reference [45] is this same manuscript (arXiv:2505.19501), so the citation does not provide independent support for Genome-Bench, which anchors all empirical claims in the paper. The self-reference is misleading because it implies a separate companion paper validates the benchmark. However, the present paper itself describes the full construction pipeline in Section 3 and Appendix B, and the RL results are measured on a held-out test split with an independent human-expert comparison on 35 questions. The citation is therefore redundant rather than load-bearing: removing it does not change any derivation, and no reported accuracy is equivalent to a fitted parameter or a self-cited theorem.

full rationale

The only identifiable circularity marker is the self-citation of reference [45], which shares the paper's own arXiv ID and title for Genome-Bench. This is a minor citation hygiene issue, not a load-bearing circular step, because Section 3 and Appendix B fully specify the data pipeline within the same paper. The core empirical chain is otherwise self-contained: forum posts supply the original expert answers; GPT-4-Turbo and GPT-4o only extract, rewrite, and generate distractors; the RL models are trained on a disjoint 80% training split and evaluated on the remaining 20% test split; and the human evaluation provides an external, albeit small, reference set. Concerns that the GPT-generated distractors and labels might let models exploit stylistic artifacts are legitimate benchmark-validity risks, but they are not circularity by construction, since no equation or fitted parameter makes the test-set accuracy equal to the training labels. Accordingly, the circularity score is low.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new physical or biological entities. The main unstated assumptions are the reliability of the forum ground truth and of the LLM-based data generation pipeline. The reward function contains hand-chosen weights that directly influence the training signal.

free parameters (3)
  • Answer correctness reward weight = +2
    Chosen by hand in the reward function. The central RL objective depends on this weight.
  • Format compliance reward weight = +1
    Chosen by hand, giving partial credit for a well-formatted but incorrect answer.
  • GRPO KL penalty coefficient (beta) = 0.005
    Set manually as a hyperparameter; affects how much the policy can drift from the reference model.
assumptions (3)
  • domain assumption Forum responses in the CRISPR Google Group are accurate expert ground truth.
    The benchmark labels come directly from forum replies with no external scientific validation.
  • domain assumption GPT-4-Turbo and GPT-4o faithfully extract, rewrite, and generate distractors without systematic errors.
    The entire data pipeline relies on LLM generation; human quality control is described only as high-level filtering.
  • ad hoc to paper The rule-based reward (+2 correct, +1 format) adequately captures scientific reasoning quality.
    The weights are chosen by the authors and are not derived from any principle or calibration.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Toward Scientific Reasoning in LLMs: Training from Expert Discussions via Reinforcement Learning." pith.science (2026). https://pith.science/paper/SDAS6GF2

@misc{pith2026250519501,
  author       = {Pith},
  title        = {Pith review of: Toward Scientific Reasoning in LLMs: Training from Expert Discussions via Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SDAS6GF2}},
  note         = {Machine review of arXiv:2505.19501}
}
read the original abstract

We investigate how to teach large language models (LLMs) to perform scientific reasoning by leveraging expert discussions as a learning signal. Focusing on the genomics domain, we develop an automated pipeline to extract trainable data and introduce Genome-Bench, a new benchmark constructed from over a decade of scientific forum discussions on genome engineering. Our pipeline transforms raw interactions into a reinforcement learning-friendly multiple-choice questions format, supported by 3000+ high-quality question-answer pairs spanning foundational biology, experimental troubleshooting, tool usage, and beyond. We fine-tune an LLM using RL with a rule-based reward signal derived from the synthetic MCQ dataset to enhance domain-specific reasoning. Our results show that reinforcement learning from scientific discussions improves model performance by over 15% compared to the base model on Genome-Bench, narrowing the gap between open-source LLMs and expert-level reasoning. To our knowledge, this is the first end-to-end pipeline for teaching LLMs to reason from scientific discussions, with promising potential for generalization across scientific domains beyond biology.

Figures

Figures reproduced from arXiv: 2505.19501 by the authors.

Figure 5
Figure 5. Human Evaluation and Analysis of Model Reasoning. Accuracy comparison on 35 expert￾curated questions, evaluated across Qwen2.5-7B-RL, DeepSeek-R1, the RL-Router model, and human experts. The RL-Router model achieves the highest model accuracy (67.6%), approaching the performance of human experts (73.3%). Qualitative analysis of RL-Router model-generated explanations compared to human-provided reasoning. Each example… view at source ↗
Figure 6
Figure 6. A Step-by-step Example of Genome-Bench Creation. This figure illustrates the data [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Answer distribution for Genome-Bench. Genome-Bench Evaluation Protocol To ensure consistency across all evaluations, we employed a structured system prompt that specifies the required response format. The model is instructed to first generate an internal reasoning process enclosed within <explanation>...</explanation> tags, followed by its final answer enclosed within <answer>...</answer> tags. The final answer must… view at source ↗
Figures from the paper (7 more)
Figure 8
Figure 8. Figure 8: Model accuracy across seven genome categories for commercial LLMs. [PITH_FULL_IMAGE:figures/full_fig_p020_8.png]
Figure 9
Figure 9. Figure 9: Model accuracy across seven genome categories for base and RL-fine-tuned models. [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]
Figure 10
Figure 10. Figure 10: Accuracy of models across difficulty levels. Left Commercial models in the left and base [PITH_FULL_IMAGE:figures/full_fig_p020_10.png]
Figure 11
Figure 11. Figure 11: Pass@K evaluation for RL vs Base models. All models are evaluated for K=1, 2, 4, 8, 16. The Pass@K criterion considers a question correctly answered if at least one of the K independently sampled outputs contains the correct answer. output robustness rather than intro…
Figure 12
Figure 12. Figure 12: Performance of 1.5B RL-Router model. The backbone model used for training is Qwen2.5-1.5B-Instruct model [PITH_FULL_IMAGE:figures/full_fig_p022_12.png]
Figure 13
Figure 13. Figure 13: Breakdown accuracy and routing distribution for 7B RL-Router model. [PITH_FULL_IMAGE:figures/full_fig_p022_13.png]
Figure 14
Figure 14. Figure 14: Breakdown accuracy and routing distribution of 7B RL-Router model for Base Models. [PITH_FULL_IMAGE:figures/full_fig_p023_14.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Bottom-up Domain-specific Superintelligence: A Reliable Knowledge Graph is What We Need

    cs.CL 2025-07 conditional novelty 6.0 of 10

    A language model fine-tuned on knowledge-graph-path reasoning tasks (QwQ-Med-3) beats strong baselines on a same-style benchmark but shows mixed gains on external medical QA tests.

Reference graph

Works this paper leans on

50 extracted references · 13 canonical work pages · cited by 1 Pith paper

  1. [45]

    Genome-bench: A scientific reasoning benchmark from real-world expert discussions

    Ming Yin, Yuanhao Qu, Yang Ling, Le Cong, and Wang Mengdi. Genome-bench: A scientific reasoning benchmark from real-world expert discussions. arXiv preprint arXiv:2505.19501, 2025

    arXiv 2025

  2. [1]

    https://groups

    Public google discussion group: Genome engineering crispr/cas systems. https://groups. google.com/g/crispr?pli=1

    work page

  3. [2]

    Publicly available clinical bert embeddings

    Emily Alsentzer, John R Murphy, Willie Boag, Wei-Hung Weng, Di Jin, Tristan Naumann, and Matthew McDermott. Publicly available clinical bert embeddings. arXiv preprint arXiv:1904.03323, 2019

    arXiv 1904

  4. [3]

    Scibert: A pretrained language model for scientific text

    Iz Beltagy, Kyle Lo, and Arman Cohan. Scibert: A pretrained language model for scientific text. arXiv preprint arXiv:1903.10676, 2019

    arXiv 1903

  5. [4]

    Sciassess: Benchmarking llm proficiency in scientific literature analysis

    Hengxing Cai, Xiaochen Cai, Junhan Chang, Sihang Li, Lin Yao, Changxin Wang, Zhifeng Gao, Hongshuai Wang, Yongge Li, Mujie Lin, et al. Sciassess: Benchmarking llm proficiency in scientific literature analysis. arXiv preprint arXiv:2403.01976, 2024

    arXiv 2024

  6. [5]

    Maxmin-rlhf: Alignment with diverse human preferences

    Souradip Chakraborty, Jiahao Qiu, Hui Yuan, Alec Koppel, Furong Huang, Dinesh Manocha, Amrit Singh Bedi, and Mengdi Wang. Maxmin-rlhf: Alignment with diverse human preferences. arXiv preprint arXiv:2402.08925, 2024

    arXiv 2024

  7. [6]

    Deep reinforcement learning from human preferences

    Paul F Christiano, Jan Leike, Tom B Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. In Advances in Neural Information Processing Systems, volume 30, pages 4299–4307, 2017

    work page 2017

  8. [7]

    A 5′ utr language model for decoding untranslated regions of mrna and function predictions

    Yanyi Chu, Dan Yu, Yupeng Li, Kaixuan Huang, Yue Shen, Le Cong, Jason Zhang, and Mengdi Wang. A 5′ utr language model for decoding untranslated regions of mrna and function predictions. Nature Machine Intelligence, 6(4):449–460, 2024

    work page 2024

Show all 50 references
  1. [8]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

    2025 arXiv

  2. [9]

    Temporal consistency for llm reasoning process error identification

    Jiacheng Guo, Yue Wu, Jiahao Qiu, Kaixuan Huang, Xinzhe Juan, Ling Yang, and Mengdi Wang. Temporal consistency for llm reasoning process error identification. arXiv preprint arXiv:2503.14495, 2025

    2025

  3. [10]

    Embodied llm agents learn to cooperate in organized teams

    Xudong Guo, Kaixuan Huang, Jiale Liu, Wenhui Fan, Natalia Vélez, Qingyun Wu, Huazheng Wang, Thomas L Griffiths, and Mengdi Wang. Embodied llm agents learn to cooperate in organized teams. arXiv preprint arXiv:2403.12482, 2024

    2024 arXiv

  4. [11]

    Math-perturb: Benchmarking llms’ math reasoning abilities against hard perturbations

    Kaixuan Huang, Jiacheng Guo, Zihao Li, Xiang Ji, Jiawei Ge, Wenzhe Li, Yingqing Guo, Tianle Cai, Hui Yuan, Runzhe Wang, et al. Math-perturb: Benchmarking llms’ math reasoning abilities against hard perturbations. International Conference on Machine Learning, 2025

    2025

  5. [12]

    Crispr-gpt: An llm agent for automated design of gene-editing experiments

    Kaixuan Huang, Yuanhao Qu, Ming Yin, Kanghong Zhan, Dyllan Liu, Di Yin, William A Johnson, Denny Zhou, Russ Altman, Mengdi Wang, and Le Cong. Crispr-gpt: An llm agent for automated design of gene-editing experiments. nature Biomedical Engineering, 2025

    2025

  6. [13]

    A survey on large language models for code generation

    Juyong Jiang, Fan Wang, Jiasi Shen, Sungju Kim, and Sunghun Kim. A survey on large language models for code generation. arXiv preprint arXiv:2406.00515, 2024

    2024 arXiv

  7. [14]

    What disease does this patient have? a large-scale open domain question answering dataset from medical exams

    Di Jin, Eileen Pan, Nassim Oufattole, Wei-Hung Weng, Hanyi Fang, and Peter Szolovits. What disease does this patient have? a large-scale open domain question answering dataset from medical exams. Applied Sciences, 11(14):6421, 2021. 10

    2021

  8. [15]

    Pubmedqa: A dataset for biomedical research question answering

    Qiao Jin, Bhuwan Dhingra, Zhengping Liu, William W Cohen, and Xinghua Lu. Pubmedqa: A dataset for biomedical research question answering. arXiv preprint arXiv:1909.06146, 2019

    1909 arXiv

  9. [16]

    Bioasq-qa: A manually curated corpus for biomedical question answering

    Anastasia Krithara, Anastasios Nentidis, Konstantinos Bougiatiotis, and Georgios Paliouras. Bioasq-qa: A manually curated corpus for biomedical question answering. Scientific Data, 10(1):170, 2023

    2023

  10. [17]

    Training language models to self-correct via reinforcement learning

    Aviral Kumar, Vincent Zhuang, Rishabh Agarwal, Yi Su, John D Co-Reyes, Avi Singh, Kate Baumli, Shariq Iqbal, Colton Bishop, Rebecca Roelofs, et al. Training language models to self-correct via reinforcement learning. arXiv preprint arXiv:2409.12917, 2024

    2024 arXiv

  11. [18]

    T\" ulu 3: Pushing frontiers in open language model post-training

    Nathan Lambert, Jacob Morrison, Valentina Pyatkin, Shengyi Huang, Hamish Ivison, Faeze Brahman, Lester James V Miranda, Alisa Liu, Nouha Dziri, Shane Lyu, et al. T\" ulu 3: Pushing frontiers in open language model post-training. arXiv preprint arXiv:2411.15124, 2024

    2024 arXiv

  12. [19]

    Lab-bench: Measuring capabilities of language models for biology research

    Jon M Laurent, Joseph D Janizek, Michael Ruzo, Michaela M Hinks, Michael J Hammer- ling, Siddharth Narayanan, Manvitha Ponnapati, Andrew D White, and Samuel G Rodriques. Lab-bench: Measuring capabilities of language models for biology research. arXiv preprint arXiv:2407.10362, 2024

    2024 arXiv

  13. [20]

    Biobert: a pre-trained biomedical language representation model for biomedical text mining

    Jinhyuk Lee, Wonjin Yoon, Sungdong Kim, Donghyeon Kim, Sunkyu Kim, Chan Ho So, and Jaewoo Kang. Biobert: a pre-trained biomedical language representation model for biomedical text mining. Bioinformatics, 36(4):1234–1240, 2020

    2020

  14. [21]

    Mapping the increasing use of llms in scientific papers

    Weixin Liang, Yaohui Zhang, Zhengxuan Wu, Haley Lepp, Wenlong Ji, Xuandong Zhao, Hancheng Cao, Sheng Liu, Siyu He, Zhi Huang, et al. Mapping the increasing use of llms in scientific papers. arXiv preprint arXiv:2404.01268, 2024

    2024 arXiv

  15. [22]

    Understanding r1-zero-like training: A critical perspective

    Zichen Liu, Changyu Chen, Wenjun Li, Penghui Qi, Tianyu Pang, Chao Du, Wee Sun Lee, and Min Lin. Understanding r1-zero-like training: A critical perspective. arXiv preprint arXiv:2503.20783, 2025

    2025 arXiv

  16. [23]

    Biogpt: generative pre-trained transformer for biomedical text generation and mining

    Renqian Luo, Liai Sun, Yingce Xia, Tao Qin, Sheng Zhang, Hoifung Poon, and Tie-Yan Liu. Biogpt: generative pre-trained transformer for biomedical text generation and mining. Briefings in bioinformatics, 23(6):bbac409, 2022

    2022

  17. [24]

    Biomedgpt: Open multimodal generative pre-trained transformer for biomedicine

    Yizhen Luo, Jiahuan Zhang, Siqi Fan, Kai Yang, Yushuai Wu, Mu Qiao, and Zaiqing Nie. Biomedgpt: Open multimodal generative pre-trained transformer for biomedicine. arXiv preprint arXiv:2308.09442, 2023

    2023 arXiv

  18. [25]

    Covid-qa: A question answering dataset for covid-19

    Timo Möller, Anthony Reina, Raghavan Jayakumar, and Malte Pietsch. Covid-qa: A question answering dataset for covid-19. In Proceedings of the 1st Workshop on NLP for COVID-19 at ACL 2020, 2020

    2020

  19. [26]

    Training language models to follow instructions with human feedback

    Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and ...

    2022

  20. [27]

    Medmcqa: A large-scale multi-subject multi-choice dataset for medical domain question answering

    Ankit Pal, Logesh Kumar Umapathi, and Malaikannan Sankarasubbu. Medmcqa: A large-scale multi-subject multi-choice dataset for medical domain question answering. In Conference on health, inference, and learning, pages 248–260. PMLR, 2022

    2022

  21. [28]

    A domain- specific next-generation large language model (llm) or chatgpt is required for biomedical engineering and research

    Soumen Pal, Manojit Bhattacharya, Sang-Soo Lee, and Chiranjib Chakraborty. A domain- specific next-generation large language model (llm) or chatgpt is required for biomedical engineering and research. Annals of biomedical engineering, 52(3):451–454, 2024

    2024

  22. [29]

    Humanity’s last exam

    Long Phan, Alice Gatti, Ziwen Han, Nathaniel Li, Josephina Hu, Hugh Zhang, Chen Bo Calvin Zhang, Mohamed Shaaban, John Ling, Sean Shi, et al. Humanity’s last exam. arXiv preprint arXiv:2501.14249, 2025. 11

    2025 arXiv

  23. [30]

    Scifive: a text-to-text transformer model for biomedical literature

    Long N Phan, James T Anibal, Hieu Tran, Shaurya Chanana, Erol Bahadroglu, Alec Peltekian, and Grégoire Altan-Bonnet. Scifive: a text-to-text transformer model for biomedical literature. arXiv preprint arXiv:2106.03598, 2021

    2021 arXiv

  24. [31]

    Omniscience: A domain-specialized llm for scientific reasoning and discovery

    Vignesh Prabhakar, Md Amirul Islam, Adam Atanas, Yao-Ting Wang, Joah Han, Aastha Jhunjhunwala, Rucha Apte, Robert Clark, Kang Xu, Zihan Wang, et al. Omniscience: A domain-specialized llm for scientific reasoning and discovery. arXiv preprint arXiv:2503.17604, 2025

    2025 arXiv

  25. [32]

    Gpqa: A graduate-level google-proof q&a benchmark

    David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R Bowman. Gpqa: A graduate-level google-proof q&a benchmark. In First Conference on Language Modeling, 2024

    2024

  26. [33]

    Proximal policy optimization algorithms

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017

    2017 arXiv

  27. [34]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024

    2024 arXiv

  28. [35]

    Reflexion: Language agents with verbal reinforcement learning.Advances in Neural Information Processing Systems, 36, 2024

    Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning.Advances in Neural Information Processing Systems, 36, 2024

    2024

  29. [36]

    R1-searcher: Incentivizing the search capability in llms via reinforcement learning

    Huatong Song, Jinhao Jiang, Yingqian Min, Jie Chen, Zhipeng Chen, Wayne Xin Zhao, Lei Fang, and Ji-Rong Wen. R1-searcher: Incentivizing the search capability in llms via reinforcement learning. arXiv preprint arXiv:2503.05592, 2025

    2025 arXiv

  30. [37]

    Galactica: A large language model for science

    Ross Taylor, Marcin Kardas, Guillem Cucurull, Thomas Scialom, Anthony Hartshorn, Elvis Saravia, Andrew Poulton, Viktor Kerkez, and Robert Stojnic. Galactica: A large language model for science. arXiv preprint arXiv:2211.09085, 2022

    2022 arXiv

  31. [38]

    A call for built-in biosecurity safeguards for generative ai tools

    Mengdi Wang, Zaixi Zhang, Amrit Singh Bedi, Alvaro Velasquez, Stephanie Guerra, Sheng Lin-Gibson, Le Cong, Yuanhao Qu, Souradip Chakraborty, Megan Blewett, et al. A call for built-in biosecurity safeguards for generative ai tools. Nature Biotechnology, pages 1–3, 2025

    2025

  32. [39]

    Math-shepherd: Verify and reinforce llms step-by-step without human annotations

    Peiyi Wang, Lei Li, Zhihong Shao, Runxin Xu, Damai Dai, Yifei Li, Deli Chen, Yu Wu, and Zhifang Sui. Math-shepherd: Verify and reinforce llms step-by-step without human annotations. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Vol...

    2024

  33. [40]

    Pairwise proximal policy optimization: Harnessing relative feedback for llm alignment

    Tianhao Wu, Banghua Zhu, Ruoyu Zhang, Zhaojin Wen, Kannan Ramchandran, and Jiantao Jiao. Pairwise proximal policy optimization: Harnessing relative feedback for llm alignment. arXiv preprint arXiv:2310.00212, 2023

    2023 arXiv

  34. [41]

    Darwin series: Domain specific large language models for natural science

    Tong Xie, Yuwei Wan, Wei Huang, Zhenyu Yin, Yixuan Liu, Shaozhou Wang, Qingyuan Linghu, Chunyu Kit, Clara Grazian, Wenjie Zhang, et al. Darwin series: Domain specific large language models for natural science. arXiv preprint arXiv:2308.13565, 2023

    2023 arXiv

  35. [42]

    A minimalist approach to llm reasoning: From rejection sampling to reinforce

    Wei Xiong, Jiarui Yao, Yuhui Xu, Bo Pang, Lei Wang, Doyen Sahoo, Junnan Li, Nan Jiang, Tong Zhang, Caiming Xiong, and Hanze Dong. A minimalist approach to llm reasoning: From rejection sampling to reinforce. arXiv preprint arXiv:2504.11343, 2025

    2025 arXiv

  36. [43]

    Reasonflux: Hierarchical llm reasoning via scaling thought templates

    Ling Yang, Zhaochen Yu, Bin Cui, and Mengdi Wang. Reasonflux: Hierarchical llm reasoning via scaling thought templates. arXiv preprint arXiv:2502.06772, 2025

    2025 arXiv

  37. [44]

    Smith, Christopher Parisien, Colin Compas, Cheryl Martin, Mona G

    Xi Yang, Aokun Chen, Nima PourNejatian, Hoo Chang Shin, Kaleb E. Smith, Christopher Parisien, Colin Compas, Cheryl Martin, Mona G. Flores, Ying Zhang, Tanja Magoc, Christo- pher A. Harle, Gloria Lipori, Duane A. Mitchell, William R. Hogan, Elizabeth A. Shenkman, Jiang Bian, an...

    2022 arXiv

  38. [46]

    Dapo: An open-source llm reinforcement learning system at scale

    Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Tiantian Fan, Gaohong Liu, Lingjun Liu, Xin Liu, et al. Dapo: An open-source llm reinforcement learning system at scale. arXiv preprint arXiv:2503.14476, 2025

    2025 arXiv

  39. [47]

    Bio- bart: Pretraining and evaluation of a biomedical generative language model

    Hongyi Yuan, Zheng Yuan, Ruyi Gan, Jiaxing Zhang, Yutao Xie, and Sheng Yu. Bio- bart: Pretraining and evaluation of a biomedical generative language model. arXiv preprint arXiv:2204.03905, 2022

    2022 arXiv

  40. [48]

    A generalist vision–language foundation model for diverse biomedical tasks

    Kai Zhang, Rong Zhou, Eashan Adhikarla, Zhiling Yan, Yixin Liu, Jun Yu, Zhengliang Liu, Xun Chen, Brian D Davison, Hui Ren, et al. A generalist vision–language foundation model for diverse biomedical tasks. Nature Medicine, pages 1–13, 2024

    2024

  41. [49]

    Scientific large language models: A survey on biological & chemical domains

    Qiang Zhang, Keyan Ding, Tianwen Lv, Xinda Wang, Qingyu Yin, Yiwen Zhang, Jing Yu, Yuhao Wang, Xiaotong Li, Zhuoyi Xiang, et al. Scientific large language models: A survey on biological & chemical domains. ACM Computing Surveys, 57(6):1–38, 2025

    2025

  42. [50]

    Genome Engineering using CRISPR/- Cas Systems,

    Han Zhong, Zikang Shan, Guhao Feng, Wei Xiong, Xinle Cheng, Li Zhao, Di He, Jiang Bian, and Liwei Wang. Dpo meets ppo: Reinforced token optimization for rlhf. arXiv preprint arXiv:2404.18922, 2024. 13 Appendix A Detailed pipeline for Constructing the Genome-Bench Dataset Appen...

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.