An LLM is not associative

LLMs are surprisingly good at one half of the knowledge graph problem and surprisingly bad at the other.

They understand words but not topology. Well, you can replace ‘topology’ by any related concept: centrality, multi-hop reasoning, adjacency, traversal and so on.

Feed an LLM a document and ask it to pull out entities and relations; it does this remarkably well, often rivaling dedicated NER/relation-extraction pipelines, even with no fixed schema. This is the engine behind tools like my own Knwler (https://knwler.com) turning unstructured text into knowledge graph triples.

Reasoning over the graph is a different story. Once you have the graph itself ask an LLM to traverse it, find shortest paths, detect communities, or identify the most central nodes, and accuracy collapses fast as size grows. This isn’t a prompting issue, it’s structural: LLMs operate on token sequences, not adjacency structures. They have no persistent traversal state, no native notion of topology. I presume this is intrinsic to the transformers architecture of almost all LLMs. Whether a diffusion based LLM can this, I don’t know. AI can’t compute eigenvectors or solve simple graph problems (chess anyone?).

More fundamentally: LLMs don’t have an internal representation of topology or centrality at all. They don’t “know” what a hub node is, what betweenness centrality measures structurally, or how a graph’s shape constrains paths through it They can talk about these concepts fluently (they’ve read the definitions thousands of times) but cannot compute them from a given graph, because that requires maintaining and manipulating a global structural representation, which is exactly what a transformer’s local, sequential attention doesn’t do.

The curious part is, hence, that the same model that just extracted a graph from prose often can’t answer basic questions about the graph it produced (once it’s bigger than a handful of nodes). “Reading relations out of text” and “reasoning about relational structure” are nearly unrelated skills in these models, even though both look like “graph understanding” from the outside.

Why this matters: it argues for hybrid architectures. LLM for extraction and for translating natural-language questions into structured queries (Cypher, SPARQL, Gremlin), with an actual graph engine doing the topological computation, and results handed back for synthesis. The LLM is the interface, not the reasoner.

We built LLMs that can read graphs out of the world, but not models that can think in associative ways (ie. in graphs).