Efficient Teammate Adaptation with Language-assisted Progressive Intention Alignment

Zhichao Wu, Ruiqi Xue, Yi-Chen Li, Cong Guan, Jing-Wen Yang, Lei Yuan, Yang Yu

Keywords: Teammate Adaptation, Bayesian Intention Inference, Language Assisted, Multi-agent Reinforcement Learning

TL;DR: We propose TALP: a Bayesian framework that uses language for in-episode intention inference, enabling efficient adaptation to unseen MARL teammates and boosting collaboration.

Abstract: Enabling agents to collaborate effectively with diverse and previously unseen teammates remains a core challenge in multi-agent reinforcement learning (MARL), particularly in open environments. Existing approaches typically address this issue by generating a wide range of diverse teammates during training to approximate the space of potential unseen partners at deployment. However, these methods tend to focus on a single collaboration target and struggle in tasks involving multi-modal coordination equilibria, where diverse teammate intentions exist. Moreover, despite efforts to incorporate methods like teammate modeling for teammate intention inference, existing approaches rely primarily on low-level behavioral cues with limited information density, while overlooking the critical role of high-level semantic priors, such as language descriptions of teammates’ intentions, resulting in inefficient intention identification and collaboration. Addressing these challenges, we introduce **T**eammate **A**daptation with **L**anguage-assisted **P**rogressive Intention Alignment (TALP), a novel Bayesian inference-based framework for teammate adaptation. TALP is trained to be intention-aware across varied cooperative tasks. At deployment, our two-phase approach first performs unbiased Bayesian intention inference, leveraging language as prior and interaction history as evidence, and then initiates targeted cooperation based on the inferred intention, all within a single episode. Extensive experiments across diverse gridworld and Overcooked environments show that TALP not only accurately infers teammate intentions but also considerably boosts collaborative task completion efficiency, consistently outperforming existing context-based meta-learning methods and teammate modeling approaches.

Area: Coordination, Organisations, Institutions, Norms and Ethics (COINE)

Generative A I: I acknowledge that I have read and will follow this policy.

Submission Number: 585