Visual graph scaffolds outperform text for LLM reasoning, study shows

A new study by Runlin Lei, Xiaokui Xiao, and Zhewei Wei, published on arXiv, investigates whether graphs can serve as internal reasoning assistants for large language models (LLMs)—not just external knowledge sources. The team tested this on multi-hop question answering tasks, where teacher-provided reasoning traces were rewritten as graph mind maps to guide a student LLM. Their experiments revealed a clear modality gap: when graph structures were flattened into text, benefits became limited once direct answer hints were removed. Under abstract guidance, both reasoning efficiency and answer quality degraded substantially.

In contrast, visual graph guidance remained effective even without direct answer clues, and its advantage persisted after supervised fine-tuning and KL-based distillation. The findings support the claim that graphs should be studied not only as external knowledge structures for LLMs, but also as visual scaffolds for organizing reasoning. This work could lead to new training techniques that mimic human mind-mapping for complex logical tasks, potentially making LLMs more robust in structured reasoning scenarios.