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

TL;DR

The paper extends the tool paradigm for LLMs by introducing conceptual tools and a multi-persona Think-Plan-Execute framework to handle complex, multi-source and multi-strategy dialogues. By decoupling thinking, planning, and executing, TPE improves explainability, controllability, and efficiency while enabling dynamic tool usage beyond traditional functional tools. Across FoCus, CIMA, and PsyQA datasets, TPE demonstrates superior performance and robust handling of strategy transitions, with analysis highlighting the importance of internal status, retrieval planning, and in-context learning components. The work offers a versatile approach for realistic dialogue systems requiring nuanced reasoning, personalization, and therapeutic or educational strategies, and provides code and data to enable reproduction.

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 intermediate results into a coherent response. This structured approach not only enhances the explainability and controllability of responses but also reduces token redundancy. We demonstrate the effectiveness of TPE across various dialogue response generation tasks, including multi-source (FoCus) and multi-strategy interactions (CIMA and PsyQA). This reveals its potential to handle real-world dialogue interactions that require more complicated tool learning beyond just functional tools. The full code and data will be released for reproduction.

Figures (6)

• Figure 1: Two typical dialogue systems necessitating the planning capability of LLMs to 1) call different sources of knowledge; 2) call different strategies; in order to compose the final response being more personalized and helpful. The examples here are chosen from public datasets: FoCus focus and CIMA stasaski2020cima respectively. We employ color coding to denote various sources and strategies. To ensure clarity, we consistently use the same color in responses to signify their association with the respective sources or strategies.
• Figure 2: The workflow of our proposed think-plan-execute (TPE) framework, in which Thinker, Planner, and Executor are initialized using different personas with the same LLM. It's important to note that the Planner and Executor roles are undertaken by the same persona, who follows the guidelines by the Thinker under the multi-strategy scenario.
• Figure 3: Left: The effects of different numbers of retrieved results; Right: The effects of using internal status to enrich the query.
• Figure 4: The strategy distribution shifts under different situations: (a) Ground-truth; (b) ReAct with ChatGPT; (c) TPE with ChatGPT; and (d) TPE with GPT-4. It is worth noting here the Others is not the same as the others strategy defined before but stands for all others strategies combinations such as Confirmation, Question Hint Correction, and so on.
• Figure 5: Illustration of different orders of sources planning in FoCus dataset.