IC-859LLMs rely on learned priors about graph properties (cycles exist, edges are absent) rather than analyzing the specific graph structure, causing below-majority-baseline performance and extreme structure-dependent accuracy

Bahare Fatemi, Jonathan Halcrow, Bryan Perozzi

SourceTalk like a Graph: Encoding Graphs for Large Language Models

On ER graphs, 53.96% of edge-existence queries have no edge and 81.96% of cycle-check queries have a cycle, yet PaLM 62B scores below the majority baseline on both tasks. When graph structure is varied, the effect is dramatic: cycle check accuracy is 91.7% on complete graphs (which always have cycles) but only 5.9% on path graphs (which never have cycles) under zero-shot prompting. Similarly, edge existence accuracy is 60.9% on path graphs (few edges) but 19.8% on complete graphs (all edges present). The authors attribute this to the LLM having a strong prior towards graphs having cycles and a prior that two nodes are more likely to be disconnected.

Evidence
correlational
Key metric
PaLM 62B zero-shot cycle check: complete 91.7%, path 5.9%; PaLM 62B zero-shot edge existence: path 60.9%, complete 19.8%; ER graph baselines: no edge 53.96%, has cycle 81.96%
Caveat
The prior is inferred from the pattern of results; the paper does not directly measure the internal representation of the prior.
Model
PaLM 62B
Concepts
Shortcut
Related findings
IC-858, IC-860, IC-861
Extraction
automatic-extraction