ActReview: Rebuttal-Guided Training Data and Rubric Rewards for Actionable Peer Review Generation

ActReview post-trains Qwen3-8B-Base on 40K OpenReview rebuttal-derived instances with rubric rewards to produce actionable, grounded peer-review feedback, and releases a 1,000-instance human-curated benchmark (ActReview-Bench).

The paper defines Actionable Peer-review Generation as diagnostic claim generation plus revision suggestion generation. The authors build ActReview-40K from real OpenReview review-rebuttal threads, aligning reviewer weaknesses with author responses and grounding feedback in localized paper evidence. Qwen3-8B-Base is post-trained with multi-task supervised fine-tuning followed by GRPO using rubric rewards. They also release ActReview-Bench, a human-curated 1,000-instance benchmark, on which 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 but identifies a remaining gap in technical accuracy.

• ActReview-40K is built from real OpenReview review-rebuttal threads, aligning reviewer weaknesses with author responses grounded in localized paper evidence.
• Qwen3-8B-Base is post-trained via multi-task SFT followed by GRPO with rubric rewards; the arXiv version describes the rewards as candidate-aware and weakness-specific, while the Hugging Face listing describes them as weakness-specific…
• ActReview-Bench is a human-curated benchmark of 1,000 instances.
• ActReview outperforms prior specialized review-generation models on actionability and grounding and remains competitive with strong prompt-based LLMs.
• Human evaluation confirms improved revision usefulness while noting a remaining gap in technical accuracy.