CURSOR atjaunina autocomplete cilnes modeli, lai iegūtu labākus kodēšanas ieteikumus

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

The AI ​​coding platform announced the X update, saying the model now has “21% fewer recommendations than the previous model, but a 28% higher acceptance rate.”

A blog post by Cursor explained how it was done. Reinforcement learning works with rewards, where an agent goes through trial and error in an environment multiple times to maximize cumulative reward.

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

In such situations, it would simply be better for the model to not provide any recommendations at all, rather than an inaccurate recommendation that could mislead a developer working on the code.

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

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

So all accepted TAB suggestions were rewarded, while useless TAB suggestions were found, which ultimately made the system better.

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