The Reviewer's Intuition
Imagine you are an experienced peer reviewer. Reading a claim in the introduction like "our method is robust across all domains," you don't accept it at face value—you flip to the experiments to verify. Sure enough, results in one domain are clearly weak, contradicting the introduction. You flag the inconsistency in your review.
This is the core human reviewer skill: proactive investigation. Not passively reading from start to finish and summarizing, but deciding what to look at next based on what you've already read, and chasing down every red flag.
Existing AI review systems can't do this. They either prompt an LLM directly to generate a review, run a fixed multi-stage pipeline, or have multiple agents review independently. The shared flaw: the investigation path is pre-fixed and doesn't adapt to findings.
ProReviewer: Reviewing as an Investigation
ProReviewer's core innovation is modeling the review process as a Markov Decision Process (MDP):
- State: a structured Review Log—extracted claims, pending verification questions, intermediate findings
- Action: choose which section of the paper to read next, or update the log
- Transition: update the state based on the current log and newly read content
- Reward: multi-dimensional reward—syntactic validity, review completeness, content quality, score alignment
- Removing the review log: performance drops sharply—without structured memory, the agent becomes a passive "read-and-forget" reader
- Removing MDP modeling (degenerating to a fixed pipeline): performance drops—no ability to adapt investigation strategy to findings
- Counterfactual error detection: ProReviewer is especially strong at detecting cross-section inconsistencies, a direct benefit of proactive investigation
- Paper-length robustness: as papers get longer, other methods degrade while ProReviewer stays stable—the log never "forgets" earlier content
The structured review log is key. It's not simple note-taking but a dynamic workspace:
1. Claim tracking: extract key claims ("our method is 3x faster than the baseline") 2. Question generation: generate verification questions for each claim ("does the experiments section support this speed claim?") 3. Evidence gathering: search the paper for supporting or contradicting evidence 4. Cross-validation: detect inconsistencies across sections (introduction claims vs. experimental results)
The final review is generated directly from the log, so every criticism traces back to concrete evidence.
Training: First Imitate, Then Reinforce
Training happens in two stages:
Step 1: Supervised fine-tuning (SFT) on synthetic review trajectories that demonstrate the full process of "read here, spot a red flag, verify there, update the log."
Step 2: Reinforcement learning (GRPO) 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, training on ICLR 2025 and testing 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 across five dimensions—up to 39% improvement over frontier large-model approaches and 16% over the strongest fine-tuned baseline. Human evaluation consistently prefers ProReviewer, with the highest win rate in all pairwise comparisons.
Ablations: Every Component Works
Why Can 8B Beat 397B?
The result seems counterintuitive, but the logic is clear:
1. Task specialization > general capability: reviewing isn't generic Q&A; it requires specific investigation strategies. RL teaches the 8B model "when you 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 context window to "remember" earlier content. 3. Proactive investigation > passive generation: fixed pipelines "don't know what they don't know," while the MDP framework lets the agent discover blind spots.
This matches human expert intuition: the best reviewers aren't necessarily the most knowledgeable, but they're the best at chasing down red flags.
Honest Assessment
ProReviewer has only been validated on ICLR computer science papers; its effectiveness on review norms in other disciplines (biology, physics, etc.) is unknown. The 8B model's fluency may lag behind larger models, and reviews requiring broad background knowledge may be limited.
Also, RL training depends on synthetic review trajectories—their quality ceiling becomes the model's ceiling. If the synthetic trajectories never cover certain investigation strategies, the model won't learn them.
But the core contribution is methodological: reframing review from a "generation task" to an "investigation task." That perspective shift alone is valuable. Any future AI review system that still follows the "read start to finish, then summarize" path will struggle to beat the proactive 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) Fields: AI4Science, LLM Agents, Scientific Peer Review