action-guidance

The paper proposes ActGuide-RL, a method for training agentic policies in LLMs by using human action data as guidance to overcome exploration barriers in reinforcement learning without extensive supervised fine-tuning.