Graph Serializations for Reasoning LLMs: A Study Across Knowledge Extraction and Structural Reasoning

Explain Like I'm Normal

Researchers tested 15 different ways of turning graph data into text across 9 frontier models. They found that while extraction tasks are flexible, complex structural reasoning performance swings wildly depending on whether you use DOT, XML, or other formats. For founders, this means your 'context window prep' is actually a high-stakes architectural decision.