Latent representations between subjects align through MAPA

CKA · all 12 sessions: . Computed in all 384 features, independently of the selected PCA projection.

Each curve is a trial-averaged class contrast from one recording session, pooled over temporal-lobe contacts. 11 of 12 sessions are shown; the large S3 session 1 trajectory is hidden for display only. All 12 remain in the PCA fits and reported measurements. Both options pool all 12 sessions, all three tasks, and all 64 timepoints. “All layers together” fits one mean and PCA basis to 29,952 rows × 384 features across layers 0–12. “Each layer separately” fits a separate mean and basis to 2,304 rows × 384 features at each layer. The first two PCs are shown. Fits use the unscaled class contrasts, before pre-event centering; no PCA is fitted separately for individual subjects. Each curve is centered on its mean at pre-event token positions (−0.5 to 0 s). No subject alignment, trajectory amplitude normalization, smoothing, or interpolation between layers is applied. Layer 0 is the projected encoder input; layers 1–12 are residual block outputs before readout normalization.

The shape-similarity number is mean pairwise centered temporal Gram similarity in all 384 features, after averaging the two sessions per subject. It is invariant to feature rotations and uniform scale, and uses no fitted alignment. The matched-time value and a circularly shifted reference are shown together. Similarity rises overall, but not at every layer; shared evoked response shapes also contribute. These initial results support increasingly similar response geometry, rather than identical coordinates across subjects.

These are descriptive results on Neuroprobe-Lite evaluation recordings using the MAPA iid75 ViT-S/384 checkpoint at step 55,000, extracted September 9, 2026. Condition means include all selected trials, not only held-out test trials. Contextual encoder tokens can attend to later positions; pre-event centering does not imply a causal neural baseline.