Researchers have posted a new paper on Hugging Face titled ActReview: Rebuttal-Guided Training Data and Rubric Rewards for Actionable Peer Review Generation. The work addresses what the authors call Actionable Peer-review Generation, a task motivated by the growing use of large language models for pre-submission self-review. Rather than only flagging weaknesses, the system aims to guide authors toward concrete revisions.
The framework decomposes the problem into two subtasks: diagnostic claim generation and revision suggestion generation. Its central insight is that author rebuttals reveal plausible actions for addressing reviewer concerns, making them a source of latent supervision for revision-oriented feedback.
Using real review-rebuttal threads from OpenReview, the team built ActReview-40K by aligning reviewer weaknesses with author responses and grounding the resulting feedback in localized paper evidence. They then post-trained Qwen3-8B-Base with multi-task supervised fine-tuning followed by GRPO using candidate-aware, weakness-specific rubric rewards.
The paper also introduces ActReview-Bench, a human-curated benchmark of 1,000 instances for evaluating diagnostic quality and revision usefulness. According to the authors, ActReview outperforms prior specialized review-generation models on actionability and grounding while remaining competitive with strong prompt-based LLMs. Human evaluation reportedly confirms improved revision usefulness, though a gap in technical accuracy remains. Additional analyses support generalization to held-out papers and robustness across independent judges.