Can inference scaling help reviewers catch errors humans miss?

Google's Paper Assistant Tool (PAT) reframes peer review as an inference-scaling problem rather than a single-model classification problem. A zero-shot model call skims a manuscript; PAT instead ingests the full paper and spends test-time compute traversing it — checking theoretical results, validating experiments, and re-deriving dense proofs line by line. That extra compute is the mechanism, not a detail: it produces a 34% improvement over zero-shot recall on mathematical errors in the SPOT benchmark, and in pilot deployments at STOC and ICML it surfaced critical errors that had already passed human expert reviewers.

This matters because it locates review firmly on the reliable side of the assistance/autonomy line. Since Where does AI assistance become unreliable in research?, review support is exactly the kind of externally-checkable, tool-mediated task where AI is trustworthy — a proof either type-checks or it doesn't. PAT operationalizes that: line-by-line proof verification is drudgery a human reviewer needs days for, and it is precisely where a compute-scaled agent has an edge, because the bottleneck is patient exhaustiveness, not judgment.

The deeper point connects to the generation-verification asymmetry. Since Can AI verify research outputs as fast as it generates them?, the only way verification keeps pace is to make it as compute-elastic as generation — and inference scaling is what makes an automated reviewer scale with the flood of AI-assisted submissions rather than drowning in it.