Meta's ARCTIC Reframes Code Review as Critique for AI Diffs, Reporting Zero Defects From Self-Reviewed Changes Since Launch

arXiv 2607.29516·high signalXBlueskyLinkedInCopy link

Authored by a Meta-affiliated team including Nachi Nagappan and Peter Rigby, ARCTIC argues AI coding agents now generate code faster than peer review can absorb, while existing AI reviewers over-index on style and under-index on correctness, security, and performance. It adds intent prediction (0.86 F1) from conversation logs, drift detection measuring divergence between developer intent and agent output via backtranslation (QWK 0.907 against human annotators), and a code spotlight ranking regions needing scrutiny — beating the baseline AI reviewer 2.4x on quality estimation at 5x fewer tokens. In rollout, drift scores cut code misalignment by a further 5.76 points (p = 0.026) and intent prediction drew 90.2% approval.