Programs

LLM-Guided Reinforcement Learning for Interactive Environments

This research introduces LLM-Guided Reinforcement Learning (LGRL), a novel framework designed to enhance learning in interactive environments. LGRL utilizes large language models (LLMs) to break down complex, high-level objectives into a series of manageable subgoals. This approach separates the high-level planning process from the low-level execution of actions. During the training phase, the LLM dynamically generates context-aware subgoals. These subgoals are accompanied by partial rewards, offering continuous feedback to the reinforcement learning agent. This fine-grained feedback system is crucial for encouraging structured exploration and speeding up the learning convergence process. For real-time application (inference), LGRL employs a chain-of-thought strategy, allowing the LLM to adaptively adjust subgoals based on the current state of the environment. While initially demonstrated in a specific interactive setting, the LGRL method is designed to be generalizable across a broad spectrum of goal-oriented tasks. Experimental evaluations have indicated that LGRL outperforms existing baseline approaches, achieving superior success rates, greater efficiency, and faster convergence.