OEA: Online Environmental Adaptation for Task-Oriented Language Agents

Anonymous

Abstract: Recent advancements in the field of intelligent agents, especially those leveraging large language models, have been impressively substantial. However, these models still encounter significant challenges in interactive and dynamic scenarios, such as online shopping, mainly due to their lack of knowledge of the current environment. In this paper, we propose an innovative online method for environmental adaptation. The trajectories generated by large language models during task execution are utilized to update a global action-observation tree. When encountering new tasks, our method transforms the action-observation tree into text and integrates this information into the context to aid the model in solving the task. This iterative process enables the model to progressively enhance its understanding of the environment, resulting in steadily improved performance over time. Our approach obviates the need for offline fine-tuning and serves as a versatile plug-and-play solution applicable to various scenarios. In two widely-used environments, Webshop and Alfworld, our method has significantly improved performance beyond ReAct and Reflection, achieving higher accuracy and reducing the required token expenditure.

Paper Type: long

Research Area: Dialogue and Interactive Systems

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: English