TPE: Towards Better Compositional Reasoning over Conceptual Tools with Multi-persona Collaboration

Hongru WANG, Huimin WANG, Lingzhi Wang, Minda Hu, Rui Wang, Boyang XUE, Hongyuan Lu, Fei Mi, Kam-Fai Wong

Primary Area: applications to robotics, autonomy, planning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Tool Learning, Dialogue System, Large Language Models

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: A novel multi-persona framework for LLMs to plan the use of conceptual tools in the context of dialogue systems

Abstract: Large language models (LLMs) have demonstrated exceptional performance in planning the use of various **functional tools** in question-answering, such as calculators and retrievers. In this paper, we first broaden the scope of the tool, centered around **conceptual tools** in the context of dialogue systems. A **conceptual tool** specifies a cognitive concept used to help systematic or investigative thought. Such **conceptual tools** play key roles in practice, such as multiple psychological / tutoring strategies being dynamically applied in a single turn to compose helpful responses. To further enhance the reasoning and planning capability of LLMs over these **conceptual tools**, we present a multi-persona collaboration framework: Think-Plan-Execute (*TPE*), which decouples the response generation process into three roles: thinker, planner, and executor. Specifically, the *Thinker* analyzes the internal status exhibited in the dialogue context, such as user emotions and preferences, to formulate a global guideline. The *Planner* generates executable plans to call different **conceptual tools** (a.k.a, different sources or strategies), while the *Executor* assembles all intermediate results into a coherent response. This structured approach enhances response explainability and controllability, reducing token redundancy simultaneously. We demonstrate the effectiveness of *TPE* across various dialogue response generation tasks, encompassing multi-source (FoCus) and multi-strategy interactions (CIMA and PsyQA), revealing its potential to address real-world dialogue interactions with the more complicated tool learning besides only **functional tools**. Full code and data will be released for reproduction.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4406