Dense Rewards Fail to Beat Standard GRPO Under Equal Compute
An arXiv paper systematically compared dense process rewards with standard GRPO across 1668 runs under strictly compute-matched conditions, finding that dense rewards failed to outperform sparse rewards.
2026-11-11 ~ 2026-11-12 · 2 related posts
- Dense process rewards lose to standard GRPO at matched compute, 1,668-run study finds — CWRU · 2026-11-11
- Dense rewards lose to standard GRPO at matched compute, arXiv paper finds — burny_tech · 2026-11-12