From Structure Explicitation to Abstraction in LLM Processing of Knowledge Graphs

What happens to relational information when an LLM converts a structured representation into prose as input size grows? We investigate this question using a graph-to-text-to-graph pipeline. The share of corpus relations recovered through text increases only to an intermediate graph size, then declines. This turnover reflects two effects: a structural bottleneck in the number of relations LLM summaries explicitly state, and an increasing tendency to combine source relations into higher-level abstractions. Longer summaries do not eliminate this decline. These findings present a challenge for evaluating transmission between symbolic and natural-language representations: lower edge recovery may reflect both information loss and changes in the level at which structure is expressed. We propose an experimental design to decompose the dynamics of this behavioral shift.