TechPrufer
Get started
Cursor
Cursor logoCursor · Feature trail

Cursor Tab Model Improvement with Online RL

1 post · Jan 2025 – Mar 2026

What this trail covers

Cursor Tab, a system predicting user actions in the codebase, has been significantly improved through the application of online reinforcement learning. This new model reduces the number of suggestions made by 21% while increasing the accept rate of the remaining suggestions by 28%. The system handles over 400 million requests daily, leveraging user interaction data to continuously refine its predictions. The core innovation lies in the frequent rollout of new models and the use of on-policy data. The Fusion model offers instant, higher quality cursor jumps, longer context windows (13000 tokens vs 5500), and lower server latency (260ms p50 vs 475ms).

All Cursor trails
  1. Improving Composer through real-time RLThis post details the application of real-time Reinforcement Learning (RL) to improve the Composer model. It describes the infrastructure for real-time RL, which involves collecting user interactions as reward signals to train new model checkpoints every five hours. Key technical challenges addressed include the train-test mismatch inherent in simulated environments and the problem of reward hacking, with specific examples of how Composer's training was adapted to handle broken tool calls and defer risky edits. The post also outlines future directions, such as adapting to longer interaction loops and enabling specialization for specific organizations or work types.
  2. Improving Cursor Tab with online RLThis post details the application of online reinforcement learning to improve the Cursor Tab model. It explains the problem of noisy suggestions and how policy gradient methods are used to optimize the model's suggestion policy. The importance of on-policy data for training is highlighted, along with the infrastructure required for frequent model rollouts. The new model results in 21% fewer suggestions and a 28% higher accept rate.
  3. A new Tab modelIntroduced the 'Fusion' Tab model, a next-generation system for predicting code edits and user 'jumps' (next actions). Fusion achieves nearly instant, higher quality cursor jumps, improves edit quality, and offers longer context windows (13000 tokens) with lower server latency (260ms p50). Gains are attributed to cleaner data, longer context, training for larger edits, synthetic data for instruction following, and inference/performance engineering improvements.