ARCTIC Reframes AI Code Review Around Intent, Drift, and Human Attention

Summary

The paper presents ARCTIC, an AI-powered code critique system designed for software changes generated by coding agents at a scale that traditional peer review may not handle. It argues that existing AI review tools often focus on style and best practices while giving less attention to correctness, security, and performance. ARCTIC predicts the intent of a change from conversation logs and metadata, detects drift between that intent and the agent's output through backtranslation, and creates a code spotlight that ranks the parts of a diff most deserving of human scrutiny. The system's six-theme taxonomy was derived from 18,000 code reviews. In offline evaluation, intent prediction reached 0.86 F1, while drift detection achieved a quadratic weighted kappa of 0.907 against human annotators. Code spotlight produced 2.4 times the baseline reviewer's quality-estimation performance while using five times fewer tokens. During an experimental rollout, drift scores reduced code misalignment by an additional 5.76 points, with p = 0.026, and intent prediction received 90.2% approval. The authors also report that no defects had been attributed to self-reviewed diffs since launch, although the abstract does not establish a causal link between that outcome and ARCTIC.