Actionable peer review generation improves feedback for scientific papers
ActReview: Rebuttal-Guided Training Data and Rubric Rewards for Actionable Peer Review Generation
Computation and Language
Summary
Scientists often get feedback on their papers that points out problems but doesn't always help them fix those problems. The researchers created a system called ActReview that looks at real back-and-forth between reviewers and authors to learn how to suggest specific changes. This system trains a language model to diagnose issues and offer clear revision advice based on the paper's content. Tests show ActReview gives more useful feedback than previous methods, though some technical mistakes still happen. It also works well on new papers and across different evaluators.
Large language models (LLMs)Peer reviewActionable feedbackSelf-reviewRebuttalSupervised fine-tuningRubric rewardsOpenReviewRevision suggestionBenchmarking
Authors
Yiling Ma, Yilun Zhao, Sihong Wu, Ziyu Chen, Manasi Patwardhan, Arman Cohan
Abstract
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.