Meta says AI system RADAR has auto-reviewed 535,000+ code changes and landed 331,000

Meta researchers say a production system called RADAR has already automatically reviewed more than 535,000 low-risk code changes and landed more than 331,000 of them, offering a rare public look at large-scale AI automation inside one of the world’s biggest software operations. The claims appear in a new arXiv preprint posted May 28, 2026, and describe a live internal deployment rather than a lab experiment.

Meta says the push comes as AI-assisted coding has sharply increased the amount of software its engineers produce, straining traditional human review. In the paper, lead author Chris Adams and other Meta researchers and engineers write that lines of code per human-landed code change rose 105.9% year over year, while per-developer code-change volume increased 51%. “Agentic AI” accounted for more than 80% of that growth, according to the abstract, and the share of code changes receiving timely review has declined.

The system, called RADAR, short for Risk Aware Diff Auto Review, is designed to automatically handle low-risk code changes. In the abstract, the authors say: “We deployed RADAR (Risk Aware Diff Auto Review), a multi-stage funnel that classifies each diff by authorship and source type, applies eligibility gates, static heuristics, a machine-learned Diff Risk Score, LLM-based Automated Code Review, and deterministic validation before landing qualifying changes.” In plain terms, the system first sorts incoming code changes, filters out ineligible ones, estimates risk with a machine-learning model, runs a large language model to review the code, and then checks qualifying changes with fixed validation rules before they can be merged.

The headline numbers are large. According to the abstract, RADAR has reviewed more than 535,000 code changes and landed more than 331,000 of them. The paper also says that relaxing the Diff Risk Score threshold from the 25th percentile to the 50th percentile raised the approve rate to 60.31%. The safety and efficiency results are also notable, though they come from Meta’s own analysis: RADAR-reviewed code changes had a revert rate that was one-third that of non-RADAR changes, a production incident rate that was 1/50 that of non-RADAR changes, and a 35% lower median diff review wall time. The authors say they evaluated the system using telemetry, before-and-after comparisons for policy changes, and difference-in-differences analysis.

The work builds on Meta’s earlier public discussion of its Diff Risk Score, or DRS, which the company described in an Aug. 6, 2025, engineering blog post as an internal system for predicting whether a code change could trigger a production incident. “Diff Risk Score (DRS) is an AI-powered technology built at Meta that predicts the likelihood of a code change causing a production incident, also known as a SEV,” the company wrote at the time. The new paper suggests RADAR is part of a broader Meta effort to automate software workflows based on predicted risk.

Still, the findings should be read with caution. The paper is a company-authored arXiv preprint, not peer-reviewed research, and its results have not been independently validated.

Even with that caveat, the report stands out as an unusually large public case study of risk-based automated code review running in production at scale.