MTMT: Consolidating Multiple Thinking Modes to Form a Thought Tree for Strengthening LLM

TL;DR

This work addresses the difficulty of LLMs in solving complex, multi-step problems by introducing MTMT, a graph-based framework that consolidates multiple thinking modes into a thought tree. By generating sub-questions and organizing intermediate reasoning as nodes, MTMT enables System 2-like deliberation, improved traceability, and enhanced problem-solving performance. Empirical results on GPQA, TruthfulQA, and GSM8K show MTMT yielding notable gains over a strong base model, with ablations indicating the importance of key thinking modes such as decomposition and association. The approach offers a generalizable, interpretable mechanism to leverage latent knowledge in LLMs, albeit at increased prompting costs and with limitations related to prompt length and computational resources. In future work, memory modules and efficiency optimizations could further improve robustness and scalability.

Abstract

Large language models (LLMs) have shown limitations in tasks requiring complex logical reasoning and multi-step problem-solving. To address these challenges, researchers have employed carefully designed prompts and flowcharts, simulating human cognitive processes to enhance LLM performance, such as the Chain of Thought approach. In this paper, we introduce MTMT (Multi-thinking Modes Tree), a novel method that interacts with LLMs to construct a thought tree, simulating various advanced cognitive processes, including but not limited to association, counterfactual thinking, task decomposition, and comparison. By breaking down the original complex task into simpler sub-questions, MTMT facilitates easier problem-solving for LLMs, enabling more effective utilization of the latent knowledge within LLMs. We evaluate the performance of MTMT under different parameter configurations, using GPT-4o mini as the base model. Our results demonstrate that integrating multiple modes of thinking significantly enhances the ability of LLMs to handle complex tasks.

Figures (6)

• Figure 1: The overall structure of MTMT. Starting from the main question, nodes that are "uncertain" about their own answers will generate different sub-nodes. Relevant information is passed to ancestor nodes, while irrelevant or faulty nodes are marked as deactivated.
• Figure 2: Flowchart of each thinking node. Generate and answer questions, extract information, decide the next thinking mode, and transition to or generate a thinking node.
• Figure 3: The Impact of $PPT_{0}$ and $\alpha$ on $Acc$, $ANN$, and $AP$ in the TruthfulQA Dataset.
• Figure 4: The Impact of $PPT_{0}$ and $\alpha$ on $Acc$, $ANN$, and $AP$ in the GSM8K Dataset.
• Figure 5: The Impact of $PPT_{0}$ and $\alpha$ on $Acc$, $ANN$, and $AP$ in the GPQA Dataset.