Decaying Process Rewards Boost Agentic RL Performance by 45%
October 25, 2026
In multimodal agentic reinforcement learning, decaying process rewards provide a 45.4% relative gain over outcome-only GRPO for complex video editing tasks averaging 60+ turns. The research identifies specific scenarios where fine-grained credit assignment improves long-horizon planning.
HOW THIS AFFECTS YOU
●
researcherYou should consider decaying process rewards when training agents for long-horizon, multi-turn tasks.