Evaluating 15 graph representation formats across nine frontier LLMs shows that while knowledge extraction is format-insensitive, structural reasoning accuracy depends heavily on serialization. The study suggests treating graph serialization as a dynamic orchestration decision based on the model and task.
HOW THIS AFFECTS YOU
●
builderYou should dynamically choose graph serialization formats to optimize reasoning performance in your applications.