Title: Visual Graph Scaffolds for Structural Reasoning in Large Language Models

Abstract:

While graphs are frequently employed to bolster the structured reasoning capabilities of large language models (LLMs)—typically functioning as external knowledge bases accessible during inference—this study proposes an alternative perspective. We argue that the true utility of graphs for LLMs extends beyond mere information delivery to encompass the organization of the reasoning process itself. Drawing inspiration from human cognition, where graph-structured mind maps help manage complex, branching, and converging ideas, we investigate whether graphs can act as an internal mechanism for reasoning support.

To explore this, we focused on multi-hop question answering scenarios. In our setup, reasoning traces provided by a teacher were transformed into graph-based mind maps to steer a student model. Our experimental results highlight a significant modality gap. When graph structures are converted into plain text, their effectiveness diminishes considerably once direct hints toward the correct answer are eliminated. Under these abstract, text-based guidance conditions, both the quality of the answers and the efficiency of the reasoning process suffer notable declines.

Conversely, visual graph guidance maintains its efficacy even in the absence of direct answer clues. Furthermore, the benefits of this visual approach endure through supervised fine-tuning and KL-based distillation processes. These outcomes reinforce the assertion that graphs warrant investigation not merely as external repositories of knowledge, but also as visual scaffolds essential for structuring and organizing reasoning within LLMs.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC