Meta has introduced RADAR (Risk Aware Diff Auto Review), a multi-stage system designed to enhance code review efficiency amidst increasing AI-driven code production. The deployment of RADAR has resulted in a 105.9% year-over-year growth in significant lines of code per human-landed diff and a 51% increase in per-developer diff volume, with agentic AI contributing over 80% to this growth. Despite these advancements, the timely review of diffs has declined, prompting an evaluation of the system's effectiveness. RADAR employs a risk-stratified approach to classify diffs, utilizing a machine-learned Diff Risk Score and LLM-based Automated Code Review. Analysis of over 535,000 reviewed diffs indicates that relaxing the Diff Risk Score threshold significantly increased approval rates while reducing reversion and production incident rates. Overall, RADAR has achieved a 330% reduction in median time to close and a 35% decrease in median diff review wall time, demonstrating that risk-aware automation can alleviate review bottlenecks without sacrificing safety.
LLMGenerative AINLP+1
LLMGenerative AINLPAutomation
Meta Implements RADAR for Efficient Risk-Aware Code Review Automation
arXiv AI· Chris Adams, Arjun Singh Banga, Parveen Bansal et al.· Saturday, May 30, 2026
More Articles From This Day
Generative AIMultimodal+2
Generative AIMultimodalAgentsVideo Editing
Google Unveils Gemini Omni and Gemini 3.5 Models with Innovative Video Editing Capabilities
At Google I/O 2026, Google announced the launch of its latest AI models, Gemini Omni and Gemini 3.5. Gemini Omni features the ability to create and edit videos using a combination of images, audio, text, and video inputs, allowing for high-quality video generation based on real-world knowledge. Users can edit their videos through conversational prompts, making adjustments while maintaining consistency in characters and scenes. Additionally, Gemini 3.5 Flash enhances agent capabilities, delivering high performance for complex tasks and automating workflows. The models were showcased through nine demos, illustrating their innovative features and practical applications.