Cursor said they made improvements to their Tab model, or the system that suggests code changes for developers across files, using online reinforcement learning. 

The AI coding platform announced the update on X, saying the model now makes “21% fewer suggestions than the previous model while having a 28% higher accept rate for the suggestions it makes.”

A blog posted by Cursor explained how this was done. Reinforcement learning works by way of rewards, where an agent goes through trials and errors several times over in an environment to maximise cumulative rewards. 

However, the blog noted that at times, there wasn’t enough information for the agent to know what action the user was going to take; so even if the model was made technically “smarter,” it still wouldn’t know what to do. 

In such situations, it would simply be better for the model to not give any suggestions, instead of an inaccurate suggestion that could derail the developer working on the code. 

“Policy gradient methods are a general way to optimize a ‘policy’ (in this case, the Tab model) to increase a “reward,” the blog said. The reward is a number assigned to every action taken by the Tab model. 

“By using a policy gradient algorithm, we can update the policy so that it gets a higher average reward in the future,” it noted. 

So, any accepted suggestions from Tab were rewarded while unhelpful suggestions made by Tab were discouraged which eventually made the system better.

Earlier in June, Cursor launched a web app for users to manage coding agents directly from their browsers.