Cursor trains Composer 2.5 via codebase feature deletion

Using automated codebase feature deletion and test-driven evaluation represents a highly scalable path for reinforcement learning in AI coding models. It provides realistic training tasks without the high cost of manual annotation.

• –Feature deletion puzzles ground model training in real production patterns rather than synthetic toy tasks.
• –Existing unit and integration test suites serve as cheap, deterministic reward functions.
• –Injected targeted textual feedback during trajectories helps solve credit assignment problems during long coding actions.
• –Scaling task diversity 25x directly improves the agent's generalization across multi-file edits and tool calls.