ActReview: Rebuttal-Guided Training Data and Rubric Rewards for Actionable Peer Review Generation
Yiling Ma, Yilun Zhao, Sihong Wu, Ziyu Chen, Manasi Patwardhan, Arman Cohan
Yale University · University of Chicago · TCS Research
Published September 8, 2026 · on the feed September 11, 2026 · ▲ 23 · arXiv 2609.09076
The 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.
The Brief
2:36A short podcast summary of the paper. Written and voiced by AI.
Transcript
ActReview. Rebuttal-Guided Training Data and Rubric Rewards for Actionable Peer Review Generation. By Yiling Ma and five others, from Yale University, the University of Chicago, and one other institution. Published on the eighth of September. Its code repository is public and real: about twenty source files.
Authors increasingly use LLMs to review their own papers before submitting them. That creates a demand for a particular kind of feedback: not only feedback that finds weaknesses, but feedback that guides the author toward concrete revisions. The authors name this task Actionable Peer-review Generation, and they split it into two subtasks. One is generating diagnostic claims, saying what is wrong. The other is generating revision suggestions, saying what to do about it. The hard part is the second, because a review that only diagnoses leaves the author where they started.
ActReview is a post-training framework that connects diagnoses specific to a paper with revision plans that are concrete and grounded. Its central insight concerns author rebuttals. When authors answer reviewers, their rebuttals reveal plausible actions for addressing the reviewers' concerns. That makes rebuttals a source of latent supervision for feedback aimed at revision: they show, implicitly, what a useful response to a criticism looks like.
The data comes from real review and rebuttal threads on OpenReview. From them the authors build ActReview-40K, aligning each reviewer weakness with the authors' response to it, and grounding the resulting feedback in localized evidence from the paper. Qwen3-8B-Base is then post-trained in two steps: multi-task supervised fine-tuning, followed by GRPO with rubric rewards that are aware of the candidate and specific to each weakness.
For evaluation, the authors also introduce ActReview-Bench, a human-curated benchmark of one thousand instances. It measures two things: the quality of the diagnosis, and how useful the suggested revision is.
In the experiments, ActReview outperforms earlier specialized review-generation models on actionability and on grounding, while staying competitive with strong prompt-based LLMs. Human evaluation confirms that its revision suggestions are more useful, and it also reveals a remaining gap in technical accuracy. Further analyses support generalization to held-out papers, and robustness when the judges are independent of one another.
The move that makes the framework work is noticing that the supervision already exists. Every rebuttal is an author showing how a criticism could be acted on, and aligning those responses with the weaknesses they answer turns years of review threads into training data for feedback that points somewhere.
Paper Quest · 7 chapters · about 39 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Spotting a flaw is the easy half. The quest begins with the hard half: saying what to do about it.
You are building a model to review papers before submission. Current models can spot problems, but they rarely say how to fix them, so the author is left where they started.
The road ahead
- 1 · The problem ~6 min Review models can say what is wrong with a paper, but they rarely say how to fix it, and there is little data to teach them.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Author rebuttals show how real criticisms can be fixed, so they can teach a model to give concrete, grounded revision advice.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Collect review threads, align weaknesses with responses, ground them in the paper, then fine-tune and optimize with rubric rewards.
- 6 · The experiment ~6 min You’ll guess: Human evaluators compared ActReview's feedback with other systems. On which dimension did a gap still remain?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.