Exploration and Online Transfer with Behavioral Foundation Models
2026-06-29 • Artificial Intelligence
Artificial IntelligenceMachine Learning
AI summaryⓘ
The authors study a way for reinforcement learning agents to quickly adapt to new tasks without extra training, even when they don't know the reward ahead of time. They point out that previous methods needed to see the task reward in advance offline, which isn't always possible if the reward is unknown or hidden. To fix this, they suggest using the agent itself to explore and learn about the reward during interaction, treating the learning process like a bandit problem where the agent tries different actions and improves based on feedback. They provide a mathematical approach to guide exploration efficiently and test their idea in a simple setup to show it works.
zero-shot transferreinforcement learningexploration-exploitationbehavioral foundation modelsonline learningreward functionbandit problemupper confidence boundlinear reward approximation
Authors
Louis Bagot, Mathieu Lefort, Laëtitia Matignon
Abstract
Zero-shot Transfer in Reinforcement Learning (RL) aims to train an agent that can generate optimal policies for any reward function, without additional learning at transfer time, while training only on reward-free trajectories. For their generality over tasks, such models are sometimes called ``Behavioral Foundation Models'' (BFMs). While they have shown strong performances and improvements in recent years, the current framework and algorithms still assume that, during the transfer phase, the agent is informed offline about the reward (the task to solve) through a dataset of state-reward pairs, which it uses to pick the best policy to deploy. However, in practice if the reward is a black-box (e.g. direct user feedback), it is not possible to generate such a dataset: it is necessary to observe the reward through interactions with the environment. In other words, the current framework of offline transfer is not aligned with the traditional RL setting of online learning through trial-and-error, which requires exploration in order to find rewards. This paper proposes to tackle this new online transfer in zero-shot RL, with the key insight that the BFM itself can be used to generate exploration policies. We show that it is possible to frame this online learning problem in terms of a bandit-like exploration-exploitation problem. More precisely, at each step the bandit algorithm recommends a policy, the BFM executes it in the environment, which yields a reward and a new state; we repeat the process until we converge to the optimal policy. In the popular context of linear reward approximation, we derive a formulation inspired by Upper Confidence Bound and show that exploration can be achieved through the minimization of the eigenvalues of an uncertainty matrix. We evaluate qualitatively and quantitatively our framework on a simple environment to validate the concept of our method.