Meta Unveils AI-Powered Code Quality System (CQS)
Meta has revealed its Code Quality System (CQS), an advanced AI platform designed to automate and scale the code review process. The system reportedly uses AI to identify subtle anti-patterns and architectural issues, going beyond simple syntax errors. CQS is designed to be context-aware of specific projects, aiming to provide more nuanced feedback and scale best practices across large engineering organizations.
- The system is built upon two fine-tuned Llama 3 models. One model is trained to find common code quality issues, while the second is designed to provide high-quality critiques of code reviews generated by other AI models. - To train the models, Meta engineers utilized supervised fine-tuning (SFT) and offline reinforcement learning (RL). - A team of Meta researchers, including Jalaj Bhandari, Sherman Wong, and Fan Yang, detailed the system in a research paper. - CQS has been rolled out internally to more than 5000 engineers at Meta. - The system has achieved a consistent week-over-week user-reported helpfulness rate of approximately 60%. - To prevent incorrect responses and AI hallucinations, the system incorporates a layer of hand-crafted rules to filter the output before it reaches developers. - Meta is leveraging constant developer feedback from the internal deployment as a data flywheel to iteratively retrain and improve the CQS models.
