RL Training Outperforms ReAct for LLM Agents

New research indicates that reinforcement learning can train LLM agents more effectively than current methods like ReAct-style reasoning loops or human-designed workflows.