Sakana AI LLM Peer Review Catches 73% of Core-Claim Errors

Sakana AI has published a research paper titled "Beyond Imitation" detailing a novel Large Language Model (LLM)-assisted peer review system designed for error detection. Unlike typical AI reviewers evaluated on their similarity to human reviews, this system is designed to identify deliberately planted mistakes within academic papers. The research team introduced two key components: a Contradiction Benchmark and a Multi-Layered Review (MLR) system. The Contradiction Benchmark involves inserting 1,164 contradictions across 257 papers sourced from five academic venues, serving as a testbed for the MLR system's error-finding capabilities. The MLR system is capable of processing up to 10 pages of a paper's main text and operates on readily available API models, specifically Claude Sonnet 4 and Claude Haiku 3.5, without requiring GPUs or fine-tuning. The cost associated with running this system is approximately $0.47 per review.

In performance tests, Sakana AI's MLR system demonstrated the highest error detection rate among four tested systems, achieving human-aligned scores. The system successfully identified 73.43% of core-claim errors when utilizing four reviews, significantly outperforming the best baseline system, which detected only 14.81%. While the system showed strong performance in identifying planted errors, it still struggled with real retracted arXiv papers, achieving only a 16.11% exact match rate on those specific cases. The research highlights that the MLR system's effectiveness stems from its ability to read and understand the paper's content before making judgments, enabling it to find more serious errors compared to baseline methods. However, it noted that the system can still be susceptible to subtle issues like hidden prompt injection.

The Multi-Layered Review (MLR) system is an agentic AI review framework developed by Sakana AI that prioritizes comprehension of a research paper before critique. It employs a trio of agents, all running on off-the-shelf Claude models. The Appendix Agent, powered by Claude Haiku 3.5, is responsible for summarizing experimental and implementation details found in the paper's appendix. The Literature Review Agent, utilizing Claude Sonnet 4, optionally performs web searches to contextualize the paper within existing scholarly work. The primary Review Agent, also using Claude Sonnet 4, executes a three-pass prompt chain inspired by the "Three-Pass Approach" methodology. In the first pass, it generates a high-level outline of the paper. The second pass involves a detailed reading where the agent flags potential weaknesses, underlying assumptions, and identified gaps in the research. Finally, the third pass consolidates the outputs from all agents, producing a comprehensive review that includes sections on Strengths, Weaknesses, Questions, a Recommendation, a Score, and a To-Do list. The system directly processes PDF documents, ensuring that figures and equations are preserved during the review process.