ReCollab: Retrieval-Augmented LLMs for Cooperative Ad-Hoc Teammate Modeling (arxiv.org)

A new advancement in the field of AI has emerged with the introduction of ReCollab, a retrieval-augmented framework designed to enhance cooperative ad-hoc teamwork (AHT) through large language models (LLMs). Traditional methods for modeling teammate behavior often struggle with limitations in observability and interaction, relying on fixed probabilistic models that can be insufficient in dynamic environments. This innovative approach leverages LLMs to interpret behavioral traces and improve policy adaptation, demonstrating a robust ability to classify teammate types and effectively respond to novel scenarios. The significance of ReCollab lies in its use of retrieval-augmented generation (RAG) to enhance the stability of inferences with exemplar trajectories, achieving superior performance in the cooperative Overcooked environment. The framework not only improves classification accuracy but also optimizes performance across varying layouts, revealing the potential for LLMs to serve as adaptable behavioral models in coordination tasks. These findings not only contribute to the enhancement of multiagent systems but also underscore the importance of grounding retrieval techniques in the evolving landscape of AI teamwork.