TPE: Towards Better Compositional Reasoning over Conceptual Tools with Multi-persona Collaboration
Hongru Wang††{}^{\dagger}start_FLOATSUPERSCRIPT † end_FLOATSUPERSCRIPT, Huimin Wang††{}^{\dagger}start_FLOATSUPERSCRIPT † end_FLOATSUPERSCRIPT, Lingzhi Wang††{}^{\dagger}start_FLOATSUPERSCRIPT † end_FLOATSUPERSCRIPT, Minda Hu‡‡{}^{\ddagger}start_FLOATSUPERSCRIPT ‡ end_FLOATSUPERSCRIPT, Rui Wang♢♢{}^{\diamondsuit}start_FLOATSUPERSCRIPT ♢ end_FLOATSUPERSCRIPT, Boyang Xue†normal-†{}^{\dagger}start_FLOATSUPERSCRIPT † end_FLOATSUPERSCRIPT, Hongyuan Lu†normal-†{}^{\dagger}start_FLOATSUPERSCRIPT † end_FLOATSUPERSCRIPT, Fei Mi♢normal-♢{}^{\diamondsuit}start_FLOATSUPERSCRIPT ♢ end_FLOATSUPERSCRIPT, Kam-Fai Wong†*normal-†absent{}^{\dagger*}start_FLOATSUPERSCRIPT † * end_FLOATSUPERSCRIPT †normal-†{}^{\dagger}start_FLOATSUPERSCRIPT † end_FLOATSUPERSCRIPTDepartment of Systems Engineering and Engineering Management ‡normal-‡{}^{\ddagger}start_FLOATSUPERSCRIPT ‡ end_FLOATSUPERSCRIPTDepartment of Computer Science and Engineering The Chinese University of Hong Kong {hrwang,
Abstract
Large language models (LLMs) have demonstrated exceptional performance in planning the use of various functional tools, such as calculators and retrievers, particularly in question-answering tasks. In this paper, we expand the definition of these tools, centering on conceptual tools within the context of dialogue systems. A conceptual tool specifies a cognitive concept that aids systematic or investigative thought. These conceptual tools play important roles in practice, such as multiple psychological or tutoring strategies being dynamically applied in a single turn to compose helpful responses. To further enhance the reasoning and planning capability of LLMs with these conceptual tools, we introduce a multi-persona collaboration framework: Think-Plan-Execute (TPE). This framework decouples the response generation process into three distinct 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 then generates executable plans to call different conceptual tools (e.g., sources or strategies), while the Executor compiles all int
中文速览
大语言模型（LLM）在调用计算器、检索器等"功能性工具"方面已相当出色，但现实对话场景中还存在大量"概念性工具"——如心理咨询策略、辅导策略、知识来源类型等认知层面的抽象概念——LLM 如何动态规划并组合使用这些工具仍是难题。为此，研究者提出了一种多角色协作框架 Think-Plan-Execute（TPE），将对话回复的生成过程拆分为三个角色：Thinker 负责分析用户情绪、偏好等内部状态并制定全局方向，Planner 据此生成有序、可执行的概念工具调用计划，Executor 则严格按照前两者的输出组装出最终回复。在多知识来源对话（FoCus）、辅导策略对话（CIMA）和心理咨询对话（PsyQA）三个数据集上的实验表明，TPE 在回复质量、可解释性和可控性上均优于基线方法，同时减少了冗余 token。这项工作将 LLM 的工具学习从功能性工具拓展到认知概念层面，为构建能处理复杂真实对话的智能系统提供了新思路。
原文 arXiv:2309.16090;中英对照 + 大白话阅读 https://aha.fim.ai/paper/2309.16090v1