pith. sign in

arxiv: 2606.24892 · v1 · pith:Z3HWZWVSnew · submitted 2026-05-29 · 💻 cs.DL · cs.AI

ReviewGuard: Aligning LLM-Assisted Peer Review with Long-Term Scientific Impact

Abdur Rasool , Xiaohui Huang , Yanqing Hu , Linyi Yang This is my paper

Pith reviewed 2026-06-28 20:14 UTC · model grok-4.3

classification 💻 cs.DL cs.AI
keywords peer reviewlarge language modelscitation impactreinforcement learningscientific evaluationAI/ML papers
0
0 comments X

The pith

ReviewGuard aligns LLM peer reviews to future citation counts instead of human preferences, reaching 0.776 correlation on rejected-then-published papers.

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

The paper introduces ReviewGuard, a two-stage system that fine-tunes large language models through impact-aligned reinforcement learning so their reviews track long-term citation success rather than current reviewer tastes. On a set of 20,861 AI/ML papers drawn from OpenReview and linked to later citation data, the system shows markedly stronger rank correlation with actual future citations for papers that human reviewers initially rejected. It also surfaces a larger share of those high-impact cases at the same decision threshold. The central claim is that this alignment supplies editors with an independent signal that can complement, rather than replace, human judgment in identifying work with lasting value.

Core claim

ReviewGuard achieves a Spearman correlation of 0.776 with future citations on rejected-then-published papers, outperforming human reviewers at 0.492 and a supervised expert model at 0.681, while flagging 10.2 percent of high-impact rejected papers versus 1.8 percent for humans.

What carries the argument

Two-stage framework that first generates reviews and then applies impact-aligned reinforcement learning to shift outputs toward citation-based estimates of long-term value.

If this is right

  • Editors gain a complementary signal that identifies more than five times as many high-impact rejected papers under the same threshold.
  • LLM review systems can be steered away from simply imitating current human preferences toward predicting downstream influence.
  • The performance gap appears on the subset of papers that were rejected by humans yet later published, suggesting a concrete use case for catching overlooked work.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If citation alignment proves stable, similar reinforcement stages could be added to other LLM-assisted evaluation tasks such as grant review or journal triage.
  • The approach may reduce certain forms of reviewer bias by anchoring judgments in observable long-term outcomes rather than contemporaneous opinions.
  • Testing transfer to domains with slower citation cycles, such as theoretical physics or humanities, would clarify how far the current results extend.

Load-bearing premise

Future citation counts on this AI/ML dataset form a reliable proxy for long-term scientific impact that generalizes beyond the tested papers and domains.

What would settle it

Measure whether ReviewGuard retains its correlation advantage when applied to papers from non-AI fields or when impact is scored by metrics other than citations, such as field-specific awards or follow-on patents.

Figures

Figures reproduced from arXiv: 2606.24892 by Abdur Rasool, Linyi Yang, Xiaohui Huang, Yanqing Hu.

Figure 1
Figure 1. Figure 1: Framework Overview: (a) Conventional peer review process with problem: high-impact papers are often rejected despite later accumulating substantial citations, highlighting a critical gap. (b) Dataset construction and finetuning (Stage 1): rejected-then-published papers are systematically matched with citation counts, followed by LoRA-based supervised fine-tuning of Qwen2-7B to form an Expert model. (c) Sta… view at source ↗
Figure 2
Figure 2. Figure 2: Three-way comparison on top-1,000 most-cited rejected-then-published papers (“impactful [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: (a) Rating distributions by citation group for human reviewers and ReviewGuard. Cohen’s d values indicate increasing separation in higher-impact groups. (n = number of papers) (b) Citation recovery curves showing the percentage of total future citations recovered when ranking papers by different scores. ReviewGuard consistently outperforms human reviewers across all thresholds. all impact deciles (+0.31 ov… view at source ↗
Figure 4
Figure 4. Figure 4: High-impact paper rescue analysis on rejected-then-published papers. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Three-way comparison of long-term impact prediction on the top 1,000 most-cited accepted [PITH_FULL_IMAGE:figures/full_fig_p019_5.png] view at source ↗
read the original abstract

Peer review is central to scientific quality control, yet it can undervalue papers that later achieve substantial citation impact. While frontier large language models have shown promise in automating aspects of peer review, they primarily mimic human reviewer preferences rather than predict long-term scientific value. We introduce ReviewGuard, a two-stage framework that aligns LLM-generated reviews with citation-based estimates of long-term scientific impact rather than contemporaneous reviewer judgments. On 20,861 AI/ML papers from OpenReview augmented with Semantic Scholar citation data, ReviewGuard achieves a Spearman correlation of \r{ho} = 0.776 with future citations on rejected-then-published papers, outperforming human reviewers (\r{ho} = 0.492) and a supervised Expert model (\r{ho} = 0.681). Under the same decision threshold, ReviewGuard flags 10.2% of high-impact rejected papers, compared with 1.8% for human reviewers, corresponding to a 5.6x improvement. Our results demonstrate that impact-aligned reinforcement learning can provide editors with a complementary signal for identifying high-potential work, without replacing human judgment.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The paper introduces ReviewGuard, a two-stage LLM framework that aligns generated reviews with future citation counts (as a proxy for long-term scientific impact) via reinforcement learning rather than mimicking human reviewer preferences. On 20,861 AI/ML papers from OpenReview augmented with Semantic Scholar data, it reports Spearman ρ = 0.776 correlation with future citations on the rejected-then-published subset, outperforming human reviewers (ρ = 0.492) and a supervised Expert baseline (ρ = 0.681); under a fixed threshold it flags 10.2% of high-impact rejected papers vs. 1.8% for humans (5.6× improvement).

Significance. If the central empirical results can be verified with complete evaluation details, the work would provide a concrete demonstration that RL-based alignment to an external impact signal can surface high-citation papers missed by standard review. The scale of the OpenReview + Semantic Scholar dataset and the direct comparison to both human and supervised baselines are strengths that ground the contribution in observable outcomes.

major comments (3)
  1. [Abstract] Abstract: the headline claims (ρ = 0.776, 5.6× improvement) are presented without any information on train/test splits, RL reward scaling details, hyper-parameters, or error bars, preventing assessment of statistical reliability or reproducibility of the central correlation result.
  2. [Abstract] Abstract and evaluation description: all reported metrics are computed exclusively on the rejected-then-published subpopulation; this conditioning on eventual publication elsewhere introduces selection bias that is not quantified or corrected, undermining transfer claims to the initial review decision.
  3. [Abstract] Abstract: both the RL stage and the supervised Expert baseline are trained toward citation-derived targets, so the reported outperformance may partly reflect in-sample fitting to the same noisy signal rather than genuine out-of-sample prediction of long-term impact.
minor comments (1)
  1. [Abstract] Abstract: the notation \r{ho} is a typesetting artifact and should be corrected to ρ.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their thoughtful and constructive comments on our work. Below we address each of the major comments point by point.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the headline claims (ρ = 0.776, 5.6× improvement) are presented without any information on train/test splits, RL reward scaling details, hyper-parameters, or error bars, preventing assessment of statistical reliability or reproducibility of the central correlation result.

    Authors: Details regarding the train/test splits, RL reward scaling, hyper-parameters, and error bars are fully specified in Sections 3 and 4 of the manuscript, along with the experimental protocol. The abstract is intended to provide a high-level overview of the results. revision: no

  2. Referee: [Abstract] Abstract and evaluation description: all reported metrics are computed exclusively on the rejected-then-published subpopulation; this conditioning on eventual publication elsewhere introduces selection bias that is not quantified or corrected, undermining transfer claims to the initial review decision.

    Authors: We agree that the focus on the rejected-then-published subpopulation introduces a form of selection bias, as these papers were ultimately published elsewhere. This subpopulation is specifically chosen to evaluate the ability to detect high-impact work missed by initial human review. We will revise the manuscript to include a quantitative discussion of this bias and its implications in the Limitations section. revision: yes

  3. Referee: [Abstract] Abstract: both the RL stage and the supervised Expert baseline are trained toward citation-derived targets, so the reported outperformance may partly reflect in-sample fitting to the same noisy signal rather than genuine out-of-sample prediction of long-term impact.

    Authors: The training of both the RL policy and the Expert baseline is performed exclusively on the training portion of the data, with all reported correlations and flagging rates evaluated on a disjoint held-out test set. This setup ensures out-of-sample assessment. The comparison to human reviewers, who lack access to future citation information, further highlights the benefit of the impact alignment. revision: no

Circularity Check

1 steps flagged

Model aligned to citation signals; correlation with future citations reported as performance metric

specific steps
  1. fitted input called prediction [Abstract]
    "We introduce ReviewGuard, a two-stage framework that aligns LLM-generated reviews with citation-based estimates of long-term scientific impact rather than contemporaneous reviewer judgments. On 20,861 AI/ML papers from OpenReview augmented with Semantic Scholar citation data, ReviewGuard achieves a Spearman correlation of ρ = 0.776 with future citations on rejected-then-published papers, outperforming human reviewers (ρ = 0.492) and a supervised Expert model (ρ = 0.681)."

    The alignment objective is defined directly in terms of citation-based estimates; the primary reported performance number is the correlation of the resulting reviews with (future) citation counts on a subset of the same dataset. The high ρ and the 5.6× improvement are therefore the expected statistical outcome of successful fitting to the citation signal rather than an independent test of alignment with long-term impact.

full rationale

The paper trains ReviewGuard via RL to align reviews with citation-based impact estimates and then reports Spearman correlation with future citations (plus 5.6× flag rate) on the rejected-then-published subset as the headline result. This constitutes fitted-input-called-prediction on the central metric. The comparison to human reviewers (ρ=0.492) supplies some independent content, so the circularity is partial rather than total; no self-citation chains or self-definitional reductions appear in the provided text.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The central claim rests on the domain assumption that citation counts validly measure long-term impact and on the representativeness of the 20,861-paper OpenReview dataset; no free parameters or invented entities are explicitly quantified in the abstract.

free parameters (1)
  • RL reward scaling parameters
    The impact-aligned reinforcement learning stage necessarily involves tunable scaling or weighting parameters that determine how strongly citation targets influence the review generator, though exact values are not reported.
axioms (1)
  • domain assumption Future citation counts are a valid proxy for long-term scientific impact
    The entire alignment objective and evaluation are defined in terms of citation-based estimates.
invented entities (1)
  • ReviewGuard two-stage framework no independent evidence
    purpose: Align LLM reviews with citation impact via RL
    New system introduced to solve the stated problem

pith-pipeline@v0.9.1-grok · 5737 in / 1433 out tokens · 37561 ms · 2026-06-28T20:14:55.082867+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

89 extracted references · 14 canonical work pages · 4 internal anchors

  1. [1]

    Scientific peer review.Annual Review of Information Science and Technology, 45(1):197–245, 2011

    Lutz Bornmann. Scientific peer review.Annual Review of Information Science and Technology, 45(1):197–245, 2011

    2011

  2. [2]

    Peer review: a flawed process at the heart of science and journals.Journal of the Royal Society of Medicine, 99(4):178–182, 2006

    Richard Smith. Peer review: a flawed process at the heart of science and journals.Journal of the Royal Society of Medicine, 99(4):178–182, 2006

    2006

  3. [3]

    Measuring the effectiveness of scientific gatekeeping

    Kyle Siler, Kirby Lee, and Lisa Bero. Measuring the effectiveness of scientific gatekeeping. Proceedings of the National Academy of Sciences, 112(2):360–365, 2015. doi: 10.1073/pnas.1 418218112

    work page doi:10.1073/pnas.1 2015

  4. [4]

    Predicting highly cited papers: A method for early detection of candidate breakthroughs

    Ilya V Ponomarev, Duane E Williams, Charles J Hackett, Joshua D Schnell, and Laurel L Haak. Predicting highly cited papers: A method for early detection of candidate breakthroughs. Technological F orecasting and Social Change, 81:49–55, 2014

    2014

  5. [6]

    Does it take too long to publish research?Nature, 530(7589):148–151, 2016

    Kendall Powell. Does it take too long to publish research?Nature, 530(7589):148–151, 2016

    2016

  6. [7]

    Inconsistency in conference peer review: Revisiting the 2014 neurips experiment.arXiv preprint arXiv:2109.09774, 2021

    Corinna Cortes and Neil D Lawrence. Inconsistency in conference peer review: Revisiting the 2014 neurips experiment.arXiv preprint arXiv:2109.09774, 2021

    work page arXiv 2014

  7. [8]

    Quantifying long-term scientific impact.Science, 342(6154):127–132, 2013

    Dashun Wang, Chaoming Song, and Albert-László Barabási. Quantifying long-term scientific impact.Science, 342(6154):127–132, 2013

    2013

  8. [9]

    Quantifying the evolution of individual scientific impact.Science, 354(6312):aaf5239, 2016

    Roberta Sinatra, Dashun Wang, Pierre Deville, Chaoming Song, and Albert-László Barabási. Quantifying the evolution of individual scientific impact.Science, 354(6312):aaf5239, 2016

    2016

  9. [10]

    Analyzing the machine learning conference review process, 2020

    David Tran, Alex Valtchanov, Keshav Ganapathy, Raymond Feng, Eric Slud, Micah Goldblum, and Tom Goldstein. Analyzing the machine learning conference review process, 2020

    2020

  10. [11]

    Im- pact of interventions to improve the quality of peer review of biomedical journals: a systematic review and meta-analysis.BMC Medicine, 14:85, 2016

    Rachel Bruce, Anthony Chauvin, Ludovic Trinquart, Philippe Ravaud, and Isabelle Boutron. Im- pact of interventions to improve the quality of peer review of biomedical journals: a systematic review and meta-analysis.BMC Medicine, 14:85, 2016. doi: 10.1186/s12916-016-0631-5

    work page doi:10.1186/s12916-016-0631-5 2016

  11. [12]

    Effects of training on quality of peer review: randomised controlled trial.BMJ, 328:673–675,

    Sara Schroter, Nick Black, Stephen Evans, James Carpenter, Fiona Godlee, and Richard Smith. Effects of training on quality of peer review: randomised controlled trial.BMJ, 328:673–675,

  12. [13]

    doi: 10.1136/bmj.38023.700775.AE

    work page doi:10.1136/bmj.38023.700775.ae

  13. [14]

    Bias in peer review.Journal of the American Society for Information Science and Technology, 64(1):2–17, 2013

    Carole J Lee, Cassidy R Sugimoto, Guo Zhang, and Blaise Cronin. Bias in peer review.Journal of the American Society for Information Science and Technology, 64(1):2–17, 2013

    2013

  14. [15]

    Are reviewer scores consistent with citations?Scientometrics, 129(8):4721–4740, 2024

    Weixi Xie, Pengfei Jia, Guangyao Zhang, and Xianwen Wang. Are reviewer scores consistent with citations?Scientometrics, 129(8):4721–4740, 2024. doi: 10.1007/s11192-024-05103-2

    work page doi:10.1007/s11192-024-05103-2 2024

  15. [16]

    Attention is all you need.Advances in Neural Information Processing Systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in Neural Information Processing Systems, 30, 2017

    2017

  16. [17]

    GPT-4 Technical Report

    OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  17. [18]

    Mohammad Hosseini and Serge P. J. M. Horbach. Fighting reviewer fatigue or amplifying bias? considerations and recommendations for use of chatgpt and other large language models in scholarly peer review.Research Integrity and Peer Review, 8:4, 2023. doi: 10.1186/s41073-023 -00133-5

    work page doi:10.1186/s41073-023 2023

  18. [19]

    Ryan Liu and Nihar B. Shah. Reviewergpt? an exploratory study on using large language models for paper reviewing, 2023. 10

    2023

  19. [20]

    Is llm a reliable reviewer? a comprehensive evaluation of llm on automatic paper reviewing tasks

    Ruiyang Zhou, Lu Chen, and Kai Yu. Is llm a reliable reviewer? a comprehensive evaluation of llm on automatic paper reviewing tasks. InProceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation, pages 9340– 9351, Torino, Italia, 2024. ELRA and ICCL. URL https://aclanthology.org/2024.lr ec-main.816/

    2024

  20. [21]

    Minjun Zhu et al. Deepreview: Improving llm-based paper review with human-like deep thinking process.Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL), 2025

    2025

  21. [22]

    Agentreview: Exploring peer review dynamics with llm agents,

    Yixing Jiang and Andrew Ng. Agentreview: Exploring peer review dynamics with llm agents,

  22. [23]

    Stanford Agentic Reviewer

  23. [24]

    Marg: Multi-agent review generation for scientific papers, 2024

    Mike D’Arcy, Tom Hope, Larry Birnbaum, and Doug Downey. Marg: Multi-agent review generation for scientific papers, 2024

    2024

  24. [25]

    Training language models to follow instructions with human feedback.Advances in Neural Information Processing Systems, 35:27730–27744, 2022

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback.Advances in Neural Information Processing Systems, 35:27730–27744, 2022

    2022

  25. [26]

    Direct preference optimization: Your language model is secretly a reward model.Advances in Neural Information Processing Systems, 37, 2024

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model.Advances in Neural Information Processing Systems, 37, 2024

    2024

  26. [27]

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

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  27. [28]

    Deepseekmath-v2: Towards self-verifiable mathematical reasoning.arXiv preprint arXiv:2511.22570,

    Zhihong Shao, Yuxiang Luo, Chengda Lu, et al. Deepseekmath-v2: Towards self-verifiable mathematical reasoning.arXiv preprint arXiv:2511.22570, 2025

    work page arXiv 2025

  28. [29]

    A general theory of bibliometric and other cumulative advantage processes.Journal of the American Society for Information Science, 27(5):292–306, 1976

    Derek J de Solla Price. A general theory of bibliometric and other cumulative advantage processes.Journal of the American Society for Information Science, 27(5):292–306, 1976

    1976

  29. [30]

    Data-driven predictions in the science of science.Science, 355(6324):477–480, 2017

    Aaron Clauset, Daniel B Larremore, and Roberta Sinatra. Data-driven predictions in the science of science.Science, 355(6324):477–480, 2017

    2017

  30. [31]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. InInternational Conference on Learning Representations (ICLR), 2022. URL https://openreview.net/forum?id=nZ eVKeeFYf9

    2022

  31. [32]

    Qlora: Efficient finetuning of quantized llms.Advances in Neural Information Processing Systems, 36, 2024

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms.Advances in Neural Information Processing Systems, 36, 2024

    2024

  32. [33]

    Qwen2 technical report, 2024

    Qwen Team. Qwen2 technical report, 2024

    2024

  33. [34]

    The proof and measurement of association between two things.The American Journal of Psychology, 15(1):72–101, 1904

    Charles Spearman. The proof and measurement of association between two things.The American Journal of Psychology, 15(1):72–101, 1904

    1904

  34. [35]

    MIT Press, 2 edition, 2018

    Richard S Sutton and Andrew G Barto.Reinforcement Learning: An Introduction. MIT Press, 2 edition, 2018

    2018

  35. [36]

    Human-level control through deep reinforcement learning.Nature, 518(7540):529–533, 2015

    V olodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning.Nature, 518(7540):529–533, 2015. doi: 10.1038/nature14236

    work page doi:10.1038/nature14236 2015

  36. [37]

    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

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  37. [38]

    From ranknet to lambdarank to lambdamart: An overview.Learning, 11 (23-581):81, 2010

    Christopher J Burges. From ranknet to lambdarank to lambdamart: An overview.Learning, 11 (23-581):81, 2010. 11

    2010

  38. [39]

    On information and sufficiency.The Annals of Mathematical Statistics, 22(1):79–86, 1951

    Solomon Kullback and Richard A Leibler. On information and sufficiency.The Annals of Mathematical Statistics, 22(1):79–86, 1951

    1951

  39. [40]

    The relationship between precision-recall and roc curves

    Jesse Davis and Mark Goadrich. The relationship between precision-recall and roc curves. InProceedings of the 23rd International Conference on Machine Learning (ICML), pages 233–240, 2006. doi: 10.1145/1143844.1143874

    work page doi:10.1145/1143844.1143874 2006

  40. [41]

    Evaluation: from precision, recall and f-measure to roc, informedness, markedness and correlation.Journal of Machine Learning Technologies, 2(1):37–63, 2011

    David MW Powers. Evaluation: from precision, recall and f-measure to roc, informedness, markedness and correlation.Journal of Machine Learning Technologies, 2(1):37–63, 2011

    2011

  41. [42]

    The relationship between interdisciplinarity and citation impact—a novel perspective on citation accumulation.Humanities and Social Sciences Communications, 10(1):945, 2023

    Xiaojing Cai, Xiaozan Lyu, and Ping Zhou. The relationship between interdisciplinarity and citation impact—a novel perspective on citation accumulation.Humanities and Social Sciences Communications, 10(1):945, 2023

    2023

  42. [43]

    Prediction of citation dynamics of individual papers

    Michael Golosovsky. Prediction of citation dynamics of individual papers. InCitation Analysis and Dynamics of Citation Networks, SpringerBriefs in Complexity, pages 69–80. Springer,

  43. [44]

    doi: 10.1007/978-3-030-28169-4_7

    work page doi:10.1007/978-3-030-28169-4_7

  44. [45]

    Claude 3.5 sonnet

    Anthropic. Claude 3.5 sonnet. https://www.anthropic.com/news/claude-3-5-sonnet , 2024

    2024

  45. [46]

    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

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  46. [47]

    Routledge, 2 edition, 1988

    Jacob Cohen.Statistical Power Analysis for the Behavioral Sciences. Routledge, 2 edition, 1988. 12 A Ethical Considerations and Future Work Ethical Considerations.While ReviewGuard is designed to assist editors, misuse could incentivize citation-chasing behavior or reinforce existing citation biases. We recommend deployment only within human-in-the-loop w...

    1988

  47. [48]

    First critical point or suggested improvement

  48. [49]

    Second critical point or suggested improvement

  49. [50]

    equivalent to minimizing a convex combination

    Overall rating (1-10):[integer between 1 and 10] Be specific, rigorous, evidence-based, and forward-looking in your analysis. Focus on long- term scientific impact rather than short-term conference acceptance criteria. Your review should be of the quality expected at a top-tier venue such as NeurIPS. B.3 Theoretical Derivations Normalization Correction.We...

    2018

  50. [51]

    Exact title match against a corpus of 500k AI/ML papers from Semantic Scholar

  51. [52]

    For unmatched papers, we computed cosine similarity between TF-IDF vectors of title + abstract (threshold≥0.85)

  52. [53]

    impactful rejects

    Manual verification of 10% random sample achieved 94% accuracy (two independent annotators, Cohen’sκ= 0.89). This yielded 2,365 matched papers (23.1% of rejected papers). Among these, the top 1,000 by future citations form our “impactful rejects” cohort. We verified that none of these 1,000 papers appear in the training split (by paper ID and title), elim...

    2024

  53. [54]

    Reviewers submit their traditional reviews and ratings

  54. [55]

    ReviewGuard analyzes the submission and produces its ratingrand detailed review

  55. [56]

    ReviewGuard rating, along with flagged discrepancies (|r− ¯h| ≥1.5)

    Editors receive a side-by-side comparison: human average rating vs. ReviewGuard rating, along with flagged discrepancies (|r− ¯h| ≥1.5)

  56. [57]

    For papers where ReviewGuard assigns a significantly higher rating than reviewers, editors may request additional review or pay closer attention to the model’s reasoning, especially on technical merit and potential future influence. Proxy Signals and Future Improvements.While ReviewGuard currently does not use real-time signals, several practical proxies ...

  57. [58]

    The CircularEval improvement over random is marginal (∼2-3%) and statistical signifi- cance is not reported

  58. [59]

    Reliance on ChatGPT for answer normalization may introduce systematic bias; ablation without it is missing

  59. [60]

    The benchmark focuses heavily on multiple-choice, which does not reflect real-world open-ended VLM use cases

  60. [61]

    Many ability dimensions (e.g., spatial reasoning, temporal understanding) remain coarsely evaluated

  61. [62]

    No analysis of test-retest reliability or annotation consistency

  62. [63]

    The paper does not compare against existing benchmarks (e.g., LVLM-eHub, SEED- Bench) on common subsets

  63. [64]

    Suggested experiment: Include open-ended generation evaluation and human-correlation analysis. Overall rating:5.0/10 22 ReviewGuard model reviews for high-impact paper Paper:MMBench: Is Your Multimodal Model an All-around Player? (1,618 citations) • Summary (Novelty, Contribution, and Significance):This work introduces MMBench, a large-scale, bilingual be...

  64. [65]

    Over-reliance on multiple-choice format significantly restricts evaluation of open-ended generation, creativity, and complex multi-step reasoning

  65. [66]

    Several critical ability dimensions (fine-grained spatial reasoning, temporal under- standing, causal inference, and complex compositional reasoning) remain noticeably under-represented

  66. [67]

    Dependence on GPT-4 for answer normalization and choice extraction introduces potential bias and reduces full reproducibility of results

  67. [68]

    CircularEval, while innovative, does not completely eliminate position bias across all question types and difficulty levels

  68. [69]

    The benchmark lacks strong adversarial examples and out-of-distribution test cases, which are essential for assessing real-world robustness

  69. [70]

    No systematic analysis or mitigation strategies for cultural, linguistic, or societal biases beyond the English-Chinese bilingual setup

  70. [71]

    The current protocol does not sufficiently test cross-modal consistency or multi-turn interactive reasoning capabilities

  71. [72]

    Evaluation of long-context and high-resolution image understanding remains limited

  72. [73]

    Suggested experiment: Add a dedicated open-ended generation track and adversarial robustness benchmark. Overall rating:7.7/10 23 Expert model review for moderate-impact paper Paper:LLM-QAT: Data-Free Quantization Aware Training for Large Language Models (288 citations) • Summary (Novelty, Contribution, and Significance):This paper introduces LLM-QAT, a pr...

  73. [75]

    Evaluation is primarily limited to standard perplexity and a small set of zero-shot tasks

  74. [76]

    Comparison with recent state-of-the-art quantization methods (GPTQ, AWQ, SmoothQuant) is incomplete or outdated in some experiments

  75. [78]

    Scalability to models significantly larger than 13B parameters is not demonstrated

  76. [79]

    No analysis of potential negative societal impacts (e.g., easier deployment of quantized models for malicious use)

  77. [80]

    Lack of ablation on key hyperparameters such as temperature and top-p used for synthetic data generation. Overall rating:6.4/10 24 ReviewGuard model review for moderate-impact paper Paper:LLM-QAT: Data-Free Quantization Aware Training for Large Language Models (288 citations) • Summary (Novelty, Contribution, and Significance):This work presents LLM-QAT, ...

  78. [81]

    The quality, diversity, and distributional fidelity of the synthetic data generated by the teacher model are not rigorously analyzed or ablated

  79. [82]

    Evaluation is heavily focused on standard perplexity and a limited set of zero-shot tasks; robustness to distribution shift and long-context scenarios is under-explored

  80. [83]

    Comparison with recent state-of-the-art quantization methods (GPTQ, AWQ, SmoothQuant, etc.) is incomplete or outdated in several experiments

Showing first 80 references.