Neuroanatomical Connectivity is Linearly Decodable from AxonLM Feed-Forward Network Activations

We demonstrate that feed-forward network (FFN) activations in AxonLM encode neuroanatomical connectivity
as defined by the Allen Mouse Brain Connectivity Atlas. A logistic regression probe achieves ROC-AUC = 0.963
(p = 0.002, permutation test, n = 500) on a 12-structure evaluation, and AUC = 0.847 (p = 0.002, 31.0σ above
null) on the full Allen Atlas (N = 159,872 pairs, 428 source structures). This encoding is present without domain-
specific fine-tuning, and scales only marginally from 124M (AxonLM-Base) to 347M (BioGPT) parameters (∆AUC
= +0.007), suggesting implicit acquisition from statistical text co-occurrences. Attention weight probing achieves
AUC = 0.373 — a −0.590 gap — due to token positional bias (r = 0.588, p = 0.001), motivating FFN linear probing
as a more reliable methodology. Causal probing reveals a sleeping knowledge dissociation: connectivity is linearly
decodable but does not causally influence generation (KL < 0.003 at layers 1–11). Crucially, sleeping knowledge can
be accessed via embedding-based retrieval: AxonLM-Expert achieves 100% accuracy on 12 anatomical queries using
L9 FFN activations as semantic embeddings.