Sakana's Multi-Layered Review: 73% of core-claim errors in a 1,164-error benchmark
From the AI digest for 2026-10-11
What happened: Sakana AI published a benchmark and system for LLM-assisted peer review (marktechpost, ~Oct 10): a 1,164-core-claim-error corpus on which Multi-Layered Review (MLR) catches 73% of errors.
What's genuinely new: The models-auditing-models genre gains its first quantified peer-review instance: a published error taxonomy plus a catch rate for review-time claims. It sits beside the record's decision-models lane (Jev, pplx-decider, the runner) and METR's observability work — same question (can a model adjudicate another's claims), different corpus (scientific review). The 73% figure is the vendor's own number on its own benchmark; unverified.
Watch: Independent replication; editors/journals adopting review-time LLMs; catch rates on adversarial vs honest errors.
Links: Marktechpost — Sakana AI's LLM Peer Review System Catches 73% of Core-Claim Errors