TL;DR
FactReview is an automated peer-review pipeline that extracts review-relevant claims from AI/ML manuscripts and audits them against external literature and actual code execution. By grounding evaluations in real-world evidence rather than just text, it significantly improves review quality and cuts human reviewer time by up to 58% while expanding claim coverage.
What Problem Does This Paper Address?
Large Language Models (LLMs) are increasingly being deployed to assist with peer review, but they typically evaluate manuscripts in isolation. If a paper claims to achieve state-of-the-art (SOTA) results or introduces a completely novel method, a standard LLM reviewer takes it at face value. It cannot verify if the code actually runs, if the reported metrics match the code's outputs, or if a closely related paper already solved the exact same problem. This results in superficial, uninformative reviews that fail to flag overclaims or reproducibility issues.
Methods
FactReview introduces an end-to-end pipeline consisting of four main stages:
- Document Parsing and Claim Extraction: Using MinerU, the system parses the PDF into structured text and uses an LLM to break down broad assertions into atomic, testable claims (e.g., specific datasets, metrics, and baselines).
- Literature Positioning and Reference Verification: It queries Semantic Scholar and OpenAlex to build a comparison matrix of related work, while using RefCopilot to check for citation hallucinations or integrity issues.
- Execution-Based Claim Verification: For papers with public code repositories, FactReview spins up a sandboxed workspace. It automatically infers tasks, sets up environments, and runs the code. Crucially, it features a Run-Review-Fix (RRF) loop that allows up to 3 rounds of minor automated repairs (like fixing dependency paths or launch arguments) without altering the core scientific logic.
- Claim Assessment & Review Synthesis: Finally, it cross-references the extracted claims with the gathered evidence, assigning each claim one of four labels: Supported, Partially supported, In conflict, or Inconclusive. It then generates a concise review linked directly to an evidence report.
Datasets
To evaluate FactReview, the authors constructed a custom benchmark of 35 machine-learning papers containing 495 human-verified reference claims.
- Development Set: Consists of 9 papers (including CompGCN) used to tune prompts, thresholds, and rules.
- Test Set: Consists of 26 paper-disjoint manuscripts (24 from OpenReview with known accept/reject outcomes, and 2 from arXiv). Out of these, 12 papers had usable code repositories for the execution-based evaluation.
Key Insights and Ideas
- Code Execution is the Ultimate Truth-Teller: Ablation studies revealed that removing execution evidence changed 17.0% of claim statuses—more than removing any other single evidence source (like literature retrieval).
- Atomic Splitting is Essential: Broad claims (e.g., "Our model outperforms all baselines across all tasks") must be broken down. A paper might have runnable code that reproduces one specific local result, but fails to support the broader generalized claim.
- Automated Code Repair is Viable: The Run-Review-Fix (RRF) loop successfully resolved common environment and dependency blockers, boosting the paper execution success rate from 55% to 65% without violating scientific integrity.
Key Results
- High Accuracy: FactReview achieved an 84.3% F1 score for claim recovery and a 94.7% status agreement rate on matched claims.
- Human-Approved Quality: In a blind controlled study, human evaluators rated FactReview 4.72/5 overall, significantly outperforming a direct LLM reviewer (3.98/5).
- Massive Time Savings: When human reviewers were assisted by FactReview's reports and visual summaries, their average review time dropped by 58% (from 50.6 minutes down to 21.3 minutes), while their coverage of major benchmark claims jumped from 87% to 99%.
Limitations and Future Work
The authors note several limitations: the evaluation test set is relatively small (26 papers) and strictly limited to the machine-learning domain. Additionally, the claim-recovery metrics depend on an author-constructed reference inventory, and the code execution is bound to a single type of sandboxed harness. Future work includes expanding the system to evaluate broader paper populations across diverse scientific domains and testing its robustness against alternative execution harnesses and LLM backends.
Summary updated