A Reviewer's Intuition
Imagine you are a senior reviewer. Reading a claim in the introduction—"our method is robust across all domains"—you don't just accept it. You flip to the experiments section to verify. Sure enough, results for one domain are clearly underwhelming, contradicting the introduction. You flag this inconsistency in your review.
This is the core competence of human reviewers: active investigation. Not passively reading from start to finish and writing a summary, but deciding what to look at next based on what has been read, and chasing down any suspicious findings.
Existing AI review systems can't do this. They either prompt a model directly to generate a review, follow a fixed multi-stage pipeline, or have multiple agents review independently. The common flaw: the investigation path is predetermined and does not adapt to findings.
ProReviewer: Turning Review into an Investigation
ProReviewer's core innovation is modeling the review process as a Markov Decision Process (MDP):
- State: a structured Review Log—recording extracted claims, pending verification questions, and intermediate findings
- Action: choosing which part of the paper to read next, or updating the log
- Transition: updating the state based on the current log and newly read content
- Reward: a multi-dimensional reward signal—syntactic validity, review completeness, content quality, and score alignment
- Removing the Review Log: performance drops significantly—without structured memory, the agent becomes a passive reader that "reads and forgets"
- Removing MDP modeling (degrading to a fixed pipeline): performance drops—the agent can no longer adapt its investigation strategy to findings
- Counterfactual error detection: ProReviewer is especially strong at detecting cross-section inconsistencies—exactly the benefit of active investigation
- Paper-length robustness: as papers get longer, other methods degrade while ProReviewer stays stable—the log system never "forgets" earlier content
The structured Review Log is the key. It is not simple note-taking but a dynamic workspace:
1. Claim tracking: extracting key claims from the paper ("our method is 3x faster than the baseline") 2. Question generation: generating verification questions for each claim ("does the experiments section support this speed claim?") 3. Evidence collection: searching the paper for supporting or refuting evidence 4. Cross-validation: discovering inconsistencies across sections (introduction claims vs. experimental results)
The final review is generated directly from the log, so every criticism can be traced to specific evidence.
Training: Imitate First, Then Reinforce
Training proceeds in two steps:
Step 1: Supervised Fine-Tuning (SFT). Trained on synthetic review trajectories that demonstrate the full process of "read here, spot a suspicious point, verify there, update the log."
Step 2: Reinforcement Learning (GRPO). The review strategy is optimized with multi-dimensional rewards, teaching the agent to adjust investigation depth per paper. Not every paper needs deep digging, but when a suspicious claim appears, the agent must know how far to chase it.
Results: A Big Win for a Small Model
The team built a 5K dataset of ICLR paper-review pairs, trained on ICLR 2025 and tested on ICLR 2026 (ensuring test data postdates the base models' knowledge cutoffs, avoiding contamination).
Automatic evaluation covers five dimensions: Grounding, Actionability, Technical Depth, Verifiability, and Score Alignment.
| System | Average Score | |--------|---------------| | Gemini-3.1-flash-lite | Low | | Qwen3.5-397B-A17B | Low | | Best fine-tuned baseline | Medium | | ProReviewer (8B) | Highest |
ProReviewer (8B backbone) ranks first on the average of all five dimensions, improving over frontier large-model approaches by up to 39% and over the strongest fine-tuned baseline by 16%.
Human evaluation consistently favors ProReviewer too, with the highest win rate across all pairwise comparisons.
Ablations: Every Component Earns Its Keep
Why Can 8B Beat 397B?
The result looks counterintuitive, but the logic is clear:
1. Task specialization > general capability: reviewing is not generic QA—it requires specific investigation strategies. RL training lets the 8B model learn "see a suspicious claim, go verify it," while the 397B model just generates passively. 2. Structured memory > implicit memory: the Review Log externalizes key information instead of relying on the model's context window to "remember" earlier content. 3. Active investigation > passive generation: fixed pipelines "don't know what they don't know," whereas the MDP framework lets the agent proactively discover blind spots.
This matches human intuition: a good reviewer isn't necessarily the most knowledgeable person, but always the best at chasing down suspicious points.
Honest Assessment
ProReviewer has only been validated on ICLR computer science papers; its performance on review norms in other disciplines (biology, physics, etc.) is unknown. The 8B model's generation fluency may lag larger models, and reviews requiring broad background knowledge may be limited.
Also, RL training relies on synthetic review trajectories, so the quality ceiling of the synthetic data caps the model. If the synthetic trajectories don't cover certain investigation strategies, the model can't learn them.
But the core contribution is methodological: reframing review from a "generation task" to an "investigation task." This shift in perspective is valuable in itself. Any future AI review system that still follows the "read from start to finish and write a summary" path will struggle to surpass the active investigation paradigm.
---
Paper: From Passive Generation to Investigation: A Proactive Scientific Peer Review Agent Code: https://github.com/UKPLab/arxiv2026-ProReviewer Authors: Haishuo Fang, Yue Feng, Iryna Gurevych (TU Darmstadt / University of Birmingham) Field: AI4Science, LLM Agents, Scientific Peer Review