Time-Aware RL Method RPG Boosts Efficiency by 85.7% Over Vanilla RL

A new paper proposes Reward-Rate Policy Gradient (RPG), which optimizes reward per unit time under a semi-MDP framework, achieving 85.7% efficiency gains over vanilla RL on the modded-nanoGPT speedrun.

2026-10-14 ~ 2026-10-14 · 2 related posts

• Reward-Rate Policy Gradient: Teaching Agentic RL to Pay for Execution Time — sherryyangML · 2026-10-14
• Time-aware RL (RPG) beats vanilla RL by 85.7% on modded NanoGPT speedrun — sherryyangML · 2026-10-14