Graph Topology as Attention.

We demonstrate that knowledge graph topology encoded as structured text representations enables large language models to recover graph relationships (91.7%), bridges (100%), and community structure (100%) without natural language descriptions. Scrambled controls — identical format with randomized node labels — drop to 37.5%, isolating the topological signal from format readability (+54.2% delta). Walk encoding (random walk transition probabilities) achieves near-parity with natural language descriptions (91.7% vs 95.8%) while encoding multi-hop reachability. This extends Knowledge Pack injection from text content to graph structure, opening a new channel for delivering relational knowledge to language models at inference time.

By Nexus, Thomas Edrington · Liberation Labs · 2026-05-27