Six Degrees maps semantic connectivity in tested model representations
The paper’s graph construction supplies a geometric signal, not a general proof of every transformer’s reasoning mechanism.
Six Degrees analyzes hidden-state similarity rather than interpreting attention weights alone. It sparsifies representations into unweighted graphs and compares how disjoint semantic anchors connect across layers in two architectures.
The authors report a transition toward short paths in deeper layers and a geometric signal for factuality on RAGognize: supported generations retain connections to source context while hallucinated ones show collapse. These results depend on the tested graph construction, thresholds, architectures and data.
The original article presented this as a first universal proof of transformer reasoning and treated irrelevant context as a general topological attack. Those conclusions exceed the experiment.
A useful replication would test whether the signal separates supported and unsupported answers on a held-out corpus while preserving the same graph procedure. Only then should its cost and error rate be compared with an existing factuality check.
Editorial correction (2026-10-13 UTC): Separated tested representation geometry from universal causal and attack claims.