Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learn to Think: Bootstrapping LLM Reasoning Capability Through Graph Representation Learning

Hang Gao, Chenhao Zhang, Tie Wang, Junsuo Zhao, Fengge Wu, Changwen Zheng, Huaping Liu

TL;DR

Learn to Think (L2T) introduces a graph-based framework to enable adaptive reasoning in LLMs without task-specific prompts or fine-tuning. By representing the reasoning process as a graph and applying a lightweight GNN to select reasoning modes, L2T achieves real-time adjustments to both prompts and Strategy via reinforcement learning (PPO). The approach demonstrates strong gains across Sudoku, Game of 24, TruthQuest, and Creative Writing tasks, with efficiency advantages and reduced reliance on handcrafted prompts. These results suggest a scalable path toward generalizable, robust LLM reasoning across diverse problem domains.

Abstract

Large Language Models (LLMs) have achieved remarkable success across various domains. However, they still face significant challenges, including high computational costs for training and limitations in solving complex reasoning problems. Although existing methods have extended the reasoning capabilities of LLMs through structured paradigms, these approaches often rely on task-specific prompts and predefined reasoning processes, which constrain their flexibility and generalizability. To address these limitations, we propose a novel framework that leverages graph learning to enable more flexible and adaptive reasoning capabilities for LLMs. Specifically, this approach models the reasoning process of a problem as a graph and employs LLM-based graph learning to guide the adaptive generation of each reasoning step. To further enhance the adaptability of the model, we introduce a Graph Neural Network (GNN) module to perform representation learning on the generated reasoning process, enabling real-time adjustments to both the model and the prompt. Experimental results demonstrate that this method significantly improves reasoning performance across multiple tasks without requiring additional training or task-specific prompt design. Code can be found in https://github.com/zch65458525/L2T.

Learn to Think: Bootstrapping LLM Reasoning Capability Through Graph Representation Learning

TL;DR

Learn to Think (L2T) introduces a graph-based framework to enable adaptive reasoning in LLMs without task-specific prompts or fine-tuning. By representing the reasoning process as a graph and applying a lightweight GNN to select reasoning modes, L2T achieves real-time adjustments to both prompts and Strategy via reinforcement learning (PPO). The approach demonstrates strong gains across Sudoku, Game of 24, TruthQuest, and Creative Writing tasks, with efficiency advantages and reduced reliance on handcrafted prompts. These results suggest a scalable path toward generalizable, robust LLM reasoning across diverse problem domains.

Abstract

Large Language Models (LLMs) have achieved remarkable success across various domains. However, they still face significant challenges, including high computational costs for training and limitations in solving complex reasoning problems. Although existing methods have extended the reasoning capabilities of LLMs through structured paradigms, these approaches often rely on task-specific prompts and predefined reasoning processes, which constrain their flexibility and generalizability. To address these limitations, we propose a novel framework that leverages graph learning to enable more flexible and adaptive reasoning capabilities for LLMs. Specifically, this approach models the reasoning process of a problem as a graph and employs LLM-based graph learning to guide the adaptive generation of each reasoning step. To further enhance the adaptability of the model, we introduce a Graph Neural Network (GNN) module to perform representation learning on the generated reasoning process, enabling real-time adjustments to both the model and the prompt. Experimental results demonstrate that this method significantly improves reasoning performance across multiple tasks without requiring additional training or task-specific prompt design. Code can be found in https://github.com/zch65458525/L2T.
Paper Structure (53 sections, 16 equations, 4 figures, 7 tables)

This paper contains 53 sections, 16 equations, 4 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Prompt engineering.
  4. Logic and reasoning within LLM prompting.
  5. Method
  6. Reasoning Process Graph
  7. Thought Generation Framework
  8. First Step
  9. $k$-th Step
  10. (1) Reasoning process graph node classification.
  11. (2) GNN-based reasoning mode selection.
  12. (3) Thought generation.
  13. Final Step
  14. Update
  15. Experiments
  16. ...and 38 more sections

Figures (4)

  • Figure 1: A comparison between our method and conventional methods.
  • Figure 2: An example of the reasoning process graph. Each box contains a thought generated by the LLM, representing a node in the reasoning process graph. The green boxes in the graph indicate the nodes currently being processed. We classify these nodes and used their categories to guide the LLM's next steps.
  • Figure 3: The framework of the proposed method. All LLM modules uniformly utilize the same LLM.
  • Figure 4: The temperature and top-$p$ value within the reasoning process.