TL;DR of the paper, generated automatically.

Dense step rewards trail standard GRPO at matched compute.

• Monte-Carlo reward: 14.3 points behind GRPO at equal tokens, Qwen2.5-1.5B.
• Prompt picks: predicted success near 1/2 gains 0.8 to 1.7 points.
• Learning rate: GRPO at twice the initial rate comes within half a point of those gains.