Google's PAT tool automates scientific peer review with 34% better error detection

The rapid acceleration of AI-assisted science has created a systemic challenge: traditional human peer review cannot keep up with the influx of AI-generated research. To address this, researchers at Google have developed the Paper Assistant Tool (PAT), an agentic AI framework designed for deep scientific review and verification. PAT ingests full manuscripts and produces comprehensive evaluations, checking theoretical results, validating experimental designs, suggesting improvements, and identifying potential flaws. The tool leverages inference scaling techniques to tackle deeper issues than a single model call, enabling it to spot subtle errors that might otherwise slip through. This represents a step toward automating parts of the review process while keeping human evaluators in the loop.

In tests on the SPOT benchmark, PAT achieved a 34% improvement over zero-shot recall for mathematical errors. Google piloted PAT as a pre-submission tool for authors at two major Computer Science conferences: STOC (Symposium on Theory of Computing) and ICML (International Conference on Machine Learning). The pilots demonstrated PAT's ability to identify critical errors and suggest substantive improvements before papers reach human referees. By catching errors early, the tool eases the cognitive burden on reviewers while preserving their control over final decisions. The accompanying paper also proposes a taxonomy of four progressive levels of AI-human collaboration in scientific evaluation, framing the transition toward hybrid review systems that balance automation with human oversight.