New AI peer-review benchmark tests systems with planted contradictions

Sakana AI says its paper, accepted at TMLR, proposes AI support for human peer reviewers. Its benchmark plants contradictions in papers and uses a knowledge graph to estimate their severity. The lab’s Multi-Layered Review system outlines a paper, examines its details and weaknesses, then produces a review. Sakana AI says it caught more errors than other systems it tested, including on papers withdrawn over real mistakes.