Most automated peer review systems rely on textual manuscript content alone, leaving visual elements such as figures and external scholarly signals underutilized. We introduce REM-CTX, a reinforcement-learning system that incorporates auxiliary context into the review generation process via correspondence-aware reward functions. REM-CTX trains an 8B-parameter language model with Group Relative Policy Optimization (GRPO) and combines a multi-aspect quality reward with two correspondence rewards that explicitly encourage alignment with auxiliary context. Experiments on manuscripts across Computer, Biological, and Physical Sciences show that REM-CTX achieves the highest overall review quality among six baselines, outperforming other systems with substantially larger commercial models, and surpassing the next-best RL baseline across both quality and contextual grounding metrics. Ablation studies confirm that the two correspondence rewards are complementary: each selectively improves its targeted correspondence reward while preserving all quality dimensions, and the full model outperforms all partial variants. Analysis of training dynamics reveals that the criticism aspect is negatively correlated with other metrics during training, suggesting that future studies should group multi-dimension rewards for review generation.
Figures & tables
Figure 1: Comparison of scientific review generation models. (a) Vanilla or simple prompting-based review generation models, which rely primarily on textual manuscript content and internal model knowledge. (b) Structured, prompting-based review generation model [ 4 ] . (c) A multi-agent review generation model that incorporates agents to analyze the manuscript from different aspects [ 7 ] . (d) REMOR: A reinforcement learning-based review generation model that optimizes review quality using reward functions based the manuscript text only [ 9 ] . (e) Our proposed model, REM-CTX, which combines auxiliary context with reinforcement-learning optimization via GRPO and correspondence-aware reward functions, produces more grounded and informative peer reviews.
Figure 2: Figure Correspondence Reward Function (FCRF) and the Novelty Correspondence Reward Function (NCRF) datasets and model construction. (a) Sentences from human reviews are paired with auxiliary context (figure details or novelty assessments), and each pair is labeled by an LLM along two axes: relevance and consistency. (b) A ModernBERT-based classifier is trained on these labels to score new sentence–context pairs.
Figure 3: The results of Vanilla (Sonnet 4.5), [ 4 ] , MARG [ 7 ] , MAMORX [ 8 ] , Qwen3-8B, REMOR [ 9 ] , and REM-CTX performance across overall review quality (Total Aspect Coverage), dimension coverage, and correspondence reward functions. REM-CTX (TRC) is the score when the thinking traces are included in the evaluation. (a) Dimension and Correspondence scores across models. (b) Overall review quality scores are based on a composite of multiple aspect-specific metrics.
Figure 4: Correlation scores of dimension and correspondence scores calculated from a standardized reward value across training epochs (See Appendix B.1 ). This analysis reveals that criticism is negatively correlated with the novelty correspondence score and praise. The presentation & reporting is negatively correlated with materials & methods. The presentation & reporting is negatively correlated with the figure correspondence score.
Figure 5: REM-CTX scores across Computer Science ( n=130 ), Biological Science ( n=80 ), and Physical Science ( n=24 ). (a) Per-dimension scores. (b) Overall quality. Computer Science articles receive significantly higher quality scores than Biological Sciences ( p<0.01 ). (c) The scores for each minor discipline are within margins of error. Given that all minor disciplines have only 4 papers, except Computer Science, which has 130 papers. Ultimately, this plot suggests that REM-CTX favors all minor disciplines equally.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: The plot shows a standardized learning curve for each reward dimension. It shows that criticism is the only dimension that decreases over time. This is an interesting decision for the model to prioritize other dimensions at the cost of criticism .
Qwen3-8B
REM-CTX
⬇ ### Peer Review Summary ** Overall Evaluation :** The paper presents a novel approach to multi - modal sentiment analysis by incorporating contextual relationships among utterances using LSTM networks . ... --- ### ** Strengths :** 1. ** Comprehensive Literature Review **: The paper provides a thorough overview of existing work ... 2. ** Clear Methodology **: The proposed method is well - explained , with a structured approach to extracting both unimodal and multi - modal features ...
⬇ < think > ... 3. ** Strengths **: - The paper addresses an important problem in sentiment analysis ( modeling contextual relationships )... Overall , ... </ think > The authors propose an LSTM network that models the contextual relationships between utterances in a video for sentiment analysis . The paper is well - written and the proposed model is interesting , but the experimental results are not convincing enough to recommend acceptance . The authors compare their model with several baseline methods , but it is not clear whether these baselines are state - of - the - art methods or not ...
Appendix
Table 1: The review generated by Qwen3-8B closely mirrors the content of the thinking traces of REM-CTX and primarily consists of paper summaries accompanied by high-level strengths and weaknesses.
Automated peer review uses computational systems, especially language models, to assess research papers, generate feedback, and support editorial decisions.
The rapid rise in AI conference submissions has driven increasing exploration of large language models (LLMs) for peer review support. However, LLM-based reviewers often generate superficial, formulaic comments lacking substantive, evidence-grounded feedback. We attribute this to the underutilization of two key components of human reviewing: explicit rubrics and contextual grounding in existing work. To address this, we introduce REVIEWBENCH, a benchmark evaluating review text according to paper-specific rubrics derived from official guidelines, the paper's content, and human-written reviews. We further propose REVIEWGROUNDER, a rubric-guided, tool-integrated multi-agent framework that decomposes reviewing into drafting and grounding stages, enriching shallow drafts via targeted evidence consolidation. Experiments on REVIEWBENCH show that REVIEWGROUNDER, using a Phi-4-14B-based drafter and a GPT-OSS-120B-based grounding stage, consistently outperforms baselines with substantially stronger/larger backbones (e.g., GPT-4.1 and DeepSeek-R1-670B) in both alignment with human judgments and rubric-based review quality across 8 dimensions. The code is available here.
Zhuofeng Li, Yi Lu, Dongfu Jiang +7
Texas A&M University · University of Waterloo · UC San Diego +2
As LLMs are increasingly used for pre-submission self-review, there is growing demand for feedback that not only identifies weaknesses but also guides authors toward concrete revisions. We study this as Actionable Peer-review Generation and decompose it into two subtasks: diagnostic claim generation and revision suggestion generation. We introduce ActReview, a rebuttal-guided post-training framework that connects paper-specific diagnoses to concrete, grounded revision plans. Our central insight is that author rebuttals reveal plausible actions for addressing reviewer concerns and can therefore provide latent supervision for revision-oriented feedback. From real review-rebuttal threads on OpenReview, we construct ActReview-40K by aligning reviewer weaknesses with author responses and grounding the resulting feedback in localized paper evidence. We post-train Qwen3-8B-Base with multi-task supervised fine-tuning followed by GRPO using candidate-aware, weakness-specific rubric rewards. We also introduce ActReview-Bench, a human-curated benchmark of 1,000 instances for evaluating diagnostic quality and revision usefulness. Experiments show that ActReview outperforms prior specialized review-generation models on actionability and grounding while remaining competitive with strong prompt-based LLMs. Human evaluation confirms improved revision usefulness while revealing a remaining gap in technical accuracy, and additional analyses support generalization to held-out papers and robustness across independent judges.
Yiling Ma, Yilun Zhao, Sihong Wu +3
Yale University · University of Chicago · TCS Research
As automated research is increasingly adopted for writing papers, automated review has become a critical link between paper generation and quality control. However, existing methods largely rely on single-model feedback or simple review aggregation, failing to explicitly model expert responsibilities or collective deliberation in actual peer reviews, which limits effective cross-examination and reliable decision-making. Therefore, we formulate paper review as a computational modeling problem of peer review mechanisms and propose RoundtableReview, a deliberative multi-agent framework for enhancing automated research. RoundtableReview structures each paper into eight semantic slots and coordinates six academic and industry reviewers with complementary perspectives, retrieval sources, and specialized responsibilities. Through independent assessment, cross-examination, roundtable deliberation, dynamically weighted voting, and chair synthesis, the framework generates evidence-based and traceable feedback. Experiments on two automated research backbones show that the method improves paper quality in both backbones and outperforms baselines in automatic and human evaluations with fewer comments.