This work investigates how serializing a knowledge graph into text influences the performance of large language models (LLMs). Previous studies examined only a handful of serialization formats and used less capable models, reporting performance gaps of tens of percentage points. We selected 15 graph representation formats and evaluated them on nine frontier LLM configurations, covering ontology‑based knowledge extraction and structural graph reasoning tasks. Results show that knowledge extraction is relatively robust to the choice of serialization, with minor accuracy variations, whereas structural reasoning is highly sensitive, exhibiting significant accuracy differences across formats. The optimal serialization depends on the specific model, query type, and task, leading us to treat graph serialization as a dynamic orchestration decision that selects the most suitable format at inference time.
Review: Serialization is no longer a peripheral detail; it is a decisive factor for structural reasoning accuracy, and a dynamic selection approach can substantially improve LLM performance in real‑world scenarios.