Planning without Search: Refining Frontier LLMs with Offline Goal-Conditioned RL
This addresses the challenge of scaling planning capabilities to frontier LLMs without expensive RL fine-tuning, which is incremental but practical for API-based models.
The paper tackles the problem of enabling large language models to perform complex interactive tasks requiring long-horizon planning, such as negotiation and persuasion, by proposing a method that uses goal-conditioned value functions to guide LLM reasoning. The result is superior performance over RL fine-tuning and prompting methods on tasks like tool use, social deduction, and dialogue, while maintaining efficiency and scalability.
Large language models (LLMs) excel in tasks like question answering and dialogue, but complex tasks requiring interaction, such as negotiation and persuasion, require additional long-horizon reasoning and planning. Reinforcement learning (RL) fine-tuning can enable such planning in principle, but suffers from drawbacks that hinder scalability. In particular, multi-turn RL training incurs high memory and computational costs, which are exacerbated when training LLMs as policies. Furthermore, the largest LLMs do not expose the APIs necessary to be trained in such manner. As a result, modern methods to improve the reasoning of LLMs rely on sophisticated prompting mechanisms rather than RL fine-tuning. To remedy this, we propose a novel approach that uses goal-conditioned value functions to guide the reasoning of LLM agents, that scales even to large API-based models. These value functions predict how a task will unfold given an action, allowing the LLM agent to evaluate multiple possible outcomes, both positive and negative, to plan effectively. In addition, these value functions are trained over reasoning steps rather than full actions, to be a concise and light-weight module that facilitates decision-making in multi-turn interactions. We validate our method on tasks requiring interaction, including tool use, social deduction, and dialogue, demonstrating superior performance over both RL fine-tuning and prompting methods while maintaining efficiency and scalability.