From Chains to Graphs: Self-Structured Reasoning for General-Domain LLMs

TL;DR

This paper introduces Self-Graph Reasoning (SGR), enabling large language models to construct and utilize their own graph-structured reasoning processes, significantly improving reasoning consistency and performance across diverse question-answering benchmarks.

Contribution

The paper proposes a novel framework for LLMs to explicitly represent reasoning as graphs, filling the gap of self-structured reasoning in open-domain QA, and demonstrates substantial performance gains.

Findings

SGR improves reasoning consistency in LLMs.

SGR achieves a 17.74% gain over the base model.

Fine-tuned LLaMA-3.3-70B with SGR rivals GPT-4o.

Abstract

Large Language Models (LLMs) show strong reasoning ability in open-domain question answering, yet their reasoning processes are typically linear and often logically inconsistent. In contrast, real-world reasoning requires integrating multiple premises and solving subproblems in parallel. Existing methods, such as Chain-of-Thought (CoT), express reasoning in a linear textual form, which may appear coherent but frequently leads to inconsistent conclusions. Recent approaches rely on externally provided graphs and do not explore how LLMs can construct and use their own graph-structured reasoning, particularly in open-domain QA. To fill this gap, we novelly explore graph-structured reasoning of LLMs in general-domain question answering. We propose Self-Graph Reasoning (SGR), a framework that enables LLMs to explicitly represent their reasoning process as a structured graph before producing the final answer. We further construct a graph-structured reasoning dataset that merges multiple candidate reasoning graphs into refined graph structures for model training. Experiments on five QA benchmarks across both general and specialized domains show that SGR consistently improves reasoning consistency and yields a 17.74% gain over the base model. The LLaMA-3.3-70B model fine-tuned with SGR performs comparably to GPT-4o and surpasses Claude-3.5-Haiku, demonstrating the effectiveness of graph-structured reasoning.