Has Cursor started training models on private user data? (cursor.com)

Cursor has significantly enhanced its Tab code suggestion model by leveraging online reinforcement learning (RL), resulting in 21% fewer suggestions but a 28% higher acceptance rate. Unlike traditional approaches that rely on static datasets or periodic releases, Cursor continuously deploys updated models multiple times daily, using real-time user interaction data to refine its predictions. This dynamic feedback loop allows the model to learn more precisely when to offer suggestions, minimizing distractions caused by irrelevant or incorrect completions. The core technical innovation lies in applying policy gradient methods to optimize the model’s behavior based on defined rewards tied to user acceptance. Instead of simply filtering out bad suggestions after they’re generated, the RL-based model learns to proactively avoid low-value suggestions by maximizing a reward function that encourages accepted completions and penalizes rejected ones. This on-policy training approach demands rapid deployment pipelines to collect timely user feedback, enabling efficient iterative updates. By integrating reinforcement learning directly into the suggestion mechanism, Cursor demonstrates a promising path toward smarter, context-aware code completion that enhances developer productivity while reducing noise.