RbtAct: Rebuttal as Supervision for Actionable Review Feedback Generation

Sihong Wu; Yiling Ma; Yilun Zhao; Tiansheng Hu; Owen Jiang; Manasi Patwardhan; Arman Cohan

cs.CLMar 10, 2026

RbtAct: Rebuttal as Supervision for Actionable Review Feedback Generation

Authors: Sihong Wu, Yiling Ma, Yilun Zhao, Tiansheng Hu, Owen Jiang, Manasi Patwardhan, Arman Cohan

Organizations: Yale University · New York University · TCS Research

Abstract

Large language models (LLMs) are increasingly used across the scientific workflow, including to draft peer-review reports. However, many AI-generated reviews are superficial and insufficiently actionable, leaving authors without concrete, implementable guidance and motivating the gap this work addresses. We propose RbtAct, which targets actionable review feedback generation and places existing peer review rebuttal at the center of learning. Rebuttals show which reviewer comments led to concrete revisions or specific plans, and which were only defended. Building on this insight, we leverage rebuttal as implicit supervision to directly optimize a feedback generator for actionability. To support this objective, we propose a new task called perspective-conditioned segment-level review feedback generation, in which the model is required to produce a single focused comment based on the complete paper and a specified perspective such as experiments and writing. We also build a large dataset named RMR-75K that maps review segments to the rebuttal segments that address them, with perspective labels and impact categories that order author uptake. We then train the Llama-3.1-8B-Instruct model with supervised fine-tuning on review segments followed by preference optimization using rebuttal derived pairs. Experiments with human experts and LLM-as-a-judge show consistent gains in actionability and specificity over strong baselines while maintaining grounding and relevance.

Figures & tables

Figure 1: Overview of our construction pipeline for RMR-75K.

Statistic	Value
Total mappings	75,542
Total papers	4,825
Avg. reviewers per paper	3.44
Avg. mappings per paper	15.66
Distinct reviews	16,583
Avg. mappings per review	4.56

Table 1: Summary statistics for the Review-Rebuttal-Mapping-75k dataset.

Setting	Precision	Recall	F1	Cohen’s $\kappa$
Filtered set	0.93	0.90	0.91	0.80

Table 2: Gold-standard validation of review

\rightarrow

rebuttal span mappings.

\kappa

is inter-annotator agreement (IAA).

Panel of Figure 2 — Figure 2: Left: brief summary of perspective labels for review segments and impact categories for rebuttal segments. Right: normalized (100%) impact category composition by perspective.

Statistic	SFT-13k	DPO-22k
# pairs	13,300	21,822
# papers	4,637	4,825
Per-perspective	$1{,}900$ each	avg. 3,117 each
Avg. paper length	22,152	21,798
Avg. output length	62	Ch: 65, Re: 63
Conference	ICLR 2024	ICLR 2024

Table 3: Statistics for SFT dataset and preference dataset for DPO. The unit of length is tokens. “Ch” denotes Chosen and “Re” denotes Rejected.

Tier	Winner	Loser	#Pairs	Percentages
Large impact gap (Easy)
	CRP	DWC	9,887	45.3
	CRP	DRF	381	1.7
Subtotal			10,268	47.1
Medium impact gap
	SRP	DWC	4,969	22.8

Table 4: Difficulty-stratified breakdown of preference pairs in ReviewPref-DPO-22k . Pairs are constructed only within the same paper and perspective and use strictly ordered rebuttal-impact labels.

	Human					LLM-as-a-Judge
System	Action.	Spec.	Ground.	Rel.	Help.	Action.	Spec.	Ground.	Rel.	Help.
Ours
RbtAct (ours)	3.46	4.08	4.30	4.76	4.26	3.38	3.70	4.05	4.82	3.74
RbtAct -SFT	3.28	4.01	4.16	4.70	4.24	3.18	3.59	3.94	4.72	3.66
LLMs
GPT-5-chat	3.38	4.04	4.35	4.98	4.47	3.28	3.66	4.12	4.95	3.78

Table 5: Pointwise ratings on five quality dimensions. Left: Human. Right: LLM-as-a-judge. Higher is better.

	Actionability (Pairwise Win Rate %)
Winners	RbtAct	GPT-5	DeepSeek	Llama-3.1	DeepReviewer	MARG	Qwen-3	RbtAct -SFT	LimGen
RbtAct	–	57.1	63.8	61.9	65.7	68.6	66.7	65.7	76.2
GPT-5-chat	42.9	–	55.2	57.1	60.0	62.9	60.0	54.3	71.4
DeepSeek-V3.2	36.2	44.8	–	54.3	57.1	59.0	58.1	53.3	68.6
Llama-3.1-70B	38.1	42.9	45.7	–	52.4	55.2	56.2	51.4	66.7
DeepReviewer-14B	34.3	40.0	42.9	47.6	–	51.4	53.3	50.5	64.8

Table 6: LLM-as-a-judge pairwise win rates (%) on Actionability

Appendix figures & tables25 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 3: A mapping example of our RMR-75K dataset. The review and rebuttal are from the paper titled “Large Language Models as Tool Makers”.

Panel of Figure 4 — Figure 4: Prompt used to segment the weaknesses and questions parts of the review into segments.

Panel of Figure 5 — Figure 5: Prompt used for mapping review segments with rebuttal segments.

Perspective of Review Segment
Label set	Definition
Experiments	Concerns about experimental setup and design : missing or insufficient experiments, weak or absent ablations, unfair/weak baselines, unclear dataset descriptions or splits, problematic hyperparameter/seed choices, compute budgets, or training details that undermine empirical support.
Evaluation	How results are measured, analyzed, and interpreted : inappropriate/missing metrics, lack of statistical testing or error bars, insufficient analysis (e.g., no variance/sensitivity), unfair comparisons, or inconsistencies between reported numbers and claims.
Reproducibility	Ability to reproduce results: missing implementation details or pseudo-code, absent code/data/links, unspecified hyperparameters, unclear preprocessing, seeds, or hardware; insufficient instructions to replicate tables/figures.
Novelty	Originality and relation to prior work : overlap with existing methods, incremental contributions, unclear differentiation, missing or superficial positioning against closely related literature.
Theory	Theoretical correctness and justification : flawed or unstated assumptions, gaps in proofs, incorrect derivations, mismatches between theorems and algorithms, or theory not supporting the claimed guarantees.

Appendix

Table 7: Summary of perspective labels for review segments and impact labels for rebuttal segments in § 4.1 .

Perspective	CRP	SRP	VCR	DWC	DRF
Evaluation (11,257)	4,766 / 42.3%	903 / 8.0%	171 / 1.5%	5,249 / 46.6%	168 / 1.5%
Experiments (25,160)	12,059 / 47.9%	2,272 / 9.0%	401 / 1.6%	9,833 / 39.1%	595 / 2.4%
Novelty (8,585)	2,828 / 32.9%	872 / 10.2%	185 / 2.2%	4,578 / 53.3%	122 / 1.4%
Presentation (4,776)	2,894 / 60.6%	803 / 16.8%	256 / 5.4%	784 / 16.4%	39 / 0.8%
Reproducibility (4,402)	2,009 / 45.6%	465 / 10.6%	120 / 2.7%	1,747 / 39.7%	61 / 1.4%
Theory (12,822)	4,253 / 33.2%	1,110 / 8.7%	282 / 2.2%	6,859 / 53.5%	318 / 2.5%

Appendix

Table 8: Label distribution by perspective (counts / %).

Panel of Figure 6 — Figure 6: Prompt used to label which perspective a review segment belongs to.

Panel of Figure 7 — Figure 7: Prompt used to label which impact category a rebuttal segment belongs to.

Setting	SFT	DPO
Framework	LLaMA-Factory	LLaMA-Factory
Base model	Llama-3.1-8B-Instruct	Llama-3.1-8B-Instruct
Precision	bf16	bf16
GPU	H200 141 GB	H200 141 GB
Context limit ( cutoff_len )	32,768	32,768
Generation limit ( max_new_tokens )	512	512

Appendix

Table 9: Key training configuration for SFT and DPO. “Eff. batch” is per-device batch size

\times

accumulation steps. DeepSpeed denotes ZeRO-2; FA2 denotes FlashAttention-2.

Panel of Figure 8 — Figure 8: Prompt used to generate review segments by different LLM baselines.

Figure 9: The interface of our human expert evaluation. The page contains 2 tasks: pairwise preference and per-candidate 1–5 ratings.

Figure 10: Comparison guidelines for the “Actionability” criterion.

Figure 11: Comparison guidelines for the “Specificity” criterion.

Figure 12: Comparison guidelines for the “Groundedness” criterion.

Figure 13: Comparison guidelines for the “Relevance” criterion.

Figure 14: Comparison guidelines for the “Helpfulness” criterion.

Panel of Figure 15 — Figure 15: Prompt used for point-wise evaluation for LLM-as-a-judge.

Panel of Figure 16 — Figure 16: Prompt used for pairwise evaluation for LLM-as-a-judge on Actionability.

Figure 17: Pairwise win rates heatmaps by perspective (row beats column) of § 5.3

Model	BLEU@4	ROUGE-L sum	METEOR	chrF
RbtAct (Ours)	14.62	12.64	11.65	18.57
RbtAct -SFT	14.93	12.33	11.53	18.51
GPT-5-chat	11.17	10.11	9.96	24.90
DeepSeek-V3.2	10.19	9.76	8.49	17.78
Llama-3.1-70B	10.48	8.76	8.27	16.42
Qwen-3-32B	9.72	8.58	8.14	17.18

Appendix

Table 10: Automatic evaluation on 700 test instances. Values are reported as percentages.

Panel of Figure 18 — Figure 18: Case study comparing review feedback on Actionability from Experiment, Presentation and Evaluation perspectives.

Impact label	Major (Ev/Ex/Th/No/Re)	Minor (Wr/Pr)
CRP	25,915 (41.6%)	7,587 (57.0%)
SRP	5,622 (9.0%)	1,952 (14.7%)
VCR	1,159 (1.9%)	887 (6.7%)
DWC	28,266 (45.4%)	2,781 (20.9%)
DRF	1,264 (2.0%)	109 (0.8%)

Appendix

Table 11: Impact-label distribution by a perspective-based severity proxy. Major issues are more frequently defended (DWC), but still exhibit substantial author uptake: CRP+SRP is 50.7% for major issues and 71.6% for minor issues.

Bucket (by mean rating)	$n$	RbtAct	Baseline	$\Delta$ (ours – base)
Weak (bottom 1/3)	35	3.35	3.08	+0.27
Medium (middle 1/3)	35	3.33	3.11	+0.22
Strong (top 1/3)	35	3.46	3.27	+0.19

Appendix

Table 12: Actionability by paper-strength bucket.

System	Action.	Spec.	Ground.	Rel.	Help.
RbtAct (ours)	3.38	3.71	4.05	4.82	3.74
Retrieval	3.02	3.45	3.80	4.48	3.39

Appendix

Table 13: LLM-as-a-judge comparison with a retrieval baseline. The retrieval baseline returns the nearest training review segment under the same perspective while excluding segments from the same paper. Higher is better.

	Actionability				Helpfulness
System	Mean	Med. [IQR]	% $\geq$ 4	Ridit	Mean	Med. [IQR]	% $\geq$ 4	Ridit
Ours
RbtAct (ours)	3.46	4 [2, 5]	52	0.548	4.25	4 [2, 5]	82	0.509
RbtAct -SFT	3.28	3 [2, 4]	45	0.497	4.24	4 [1, 5]	81	0.505
LLMs
GPT-5-chat	3.38	3 [2, 5]	47	0.523	4.49	5 [3, 5]	91	0.574

Appendix

Table 14: Ordinal-aware summaries of human Likert ratings on Actionability and Helpfulness. Median, interquartile range, threshold rate, and ridit score are reported in addition to the descriptive mean. Higher is better.

Comparison	$W$	$p$	Effect	Result	$W$	$p$	Effect	Result
	Actionability				Helpfulness
RbtAct vs. RbtAct -SFT	1812.5	0.018	0.23	$+$	2148.0	0.641	0.04	n.s.
RbtAct vs. GPT-5-chat	1987.0	0.276	0.09	n.s.	1695.5	0.031	-0.19	$-$
RbtAct vs. DeepSeek-V3.2	1654.0	0.004	0.31	$+$	2186.5	0.784	-0.02	n.s.
RbtAct vs. Llama-3.1-70B	1875.5	0.031	0.20	$+$	1922.0	0.096	0.14	n.s.
RbtAct vs. Qwen-3-32B	1538.0	0.001	0.36	$+$	1830.5	0.041	0.18	$+$

Appendix

Table 15: Paired ordinal tests on human ratings. We use the Wilcoxon signed-rank test over matched evaluation instances and report rank-biserial correlation as the effect size. Positive effects favor RbtAct .

+

indicates that RbtAct is significantly better,

-

indicates that the baseline is significantly better, and n.s. indicates no significant difference.

Automated Peer Review

Automated peer review uses computational systems, especially language models, to assess research papers, generate feedback, and support editorial decisions.

See all 109 papers

ActReview: Rebuttal-Guided Training Data and Rubric Rewards for Actionable Peer Review Generation
Sep 8, 2026Yiling Ma, Yilun Zhao, Sihong Wu +3LLM Post-Training Automated Peer Review
Yale University · University of Chicago · TCS Research
Defend: Automated Rebuttals for Peer Review with Minimal Author Guidance
Mar 28, 2026Jyotsana Khatri, Manasi PatwardhanHuman-in-the-Loop AI Automated Peer Review
TCS Research
PRAIB: Peer Review AI Benchmark of Behaviour of LLM-Assisted Reviewing
May 28, 2026Krzysztof Żurawicki, Julia Farganus, Arkadiusz Gaweł +2LLM Evaluation Scholarly Peer Review
Department of Artificial Intelligence Wrocław University of Science and Technology

Sep 8, 2026J/K move · Enter open · S save

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

Yiling Ma, Yilun Zhao, Sihong Wu, Ziyu Chen+2

Yale University · University of Chicago · TCS Research

LLM Post-Training Automated Peer Review

Open paper AskPDF arXiv

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.

Abstract

Automated Peer Review

Explore similar work

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

Defend: Automated Rebuttals for Peer Review with Minimal Author Guidance

PRAIB: Peer Review AI Benchmark of Behaviour of LLM-Assisted Reviewing