Generate explorative goals with large language model guidance

Xi Yuan, YAN ZHENG, Fuzheng Zhang, Di ZHANG, Hangyu Mao, Kun Gai

Keywords: Reinforcement Learning, Large Language Models, Goal-Conditioned RL, Exploration, Model-baed Reinforcement Learning

TL;DR: A hierarchical policy that combines LLM guidance with intrinsic exploration reward by learning to propose meaningful goals.

Abstract:

Reinforcement learning (RL) struggles with sparse reward environments.
Recent developments in intrinsic motivation have revealed the potential of language models to guide agents in exploring the environment.
However, the mismatch between the granularity of environment transitions and natural language descriptions hinders effective exploration for current methods.
To address this problem, we introduce a model-based RL method named Language-Guided Explorative Goal Generation (LanGoal), which combines large language model (LLM) guidance with intrinsic exploration reward by learning to propose meaningful goals.
LanGoal learns a hierarchical policy together with a world model. The high-level policy learns to propose goals based on LLM guidance to explore the environment, and the low-level policy learns to achieve the goals.
Extensive results on Crafter demonstrate the effectiveness of LanGoal compared to recent methods.

Primary Area: reinforcement learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 13843