Cursor Composer 2: RL Algorithm Details

Source: Tweet thread Author: Unknown (ML researcher) Date: 2026-03-30

Summary

Tweet revealing key details about Cursor Composer 2’s RL training algorithm — very close to GRPO but with specific modifications.

Key Claims

  • Uses GRPO-like algorithm but removes length normalization.
  • Removes group advantage standardization.
  • Adds KL term with K1 variant (simpler KL approximation).
  • Non-linear length penalty: penalizes very long outputs more than linearly.
  • These choices suggest Cursor found standard GRPO had verbosity and instability issues that these modifications address.

Connection to Other Sources

Complements GRPO++ and Cursor Tab Online RL. Shows Cursor’s continued investment in RL for code completion.

Concepts