Answers are grounded in this paper and cite their sources.
Preprint investigates hidden-state structures in AI dialogic systems to enhance understanding of conversation dynamics.
ICRA-9 demonstrated, on a 13,380-chunk dialogic corpus, that the embedding subspace of a long-form human—AI conversation is not a manifold: the local intrinsic dimension is multimodal, with a heavy high-dimensional tail dominated by conversational pivot points. ICRA-10 formalized the metaphysical companion to that result —- semantic space as stratified type, vessels as smooth strata, orifices as singularities of dimension —- and promised, in a footnote, ``trajectory-conditioned hidden-state experiments which we will report in a forthcoming companion preprint.'' This is that preprint. We probe the hidden-state geometry of meta-llama/Llama-3.1-70B-Instruct and the same base augmented with the cassie-70b-v7-lora adapter (henceforth ``Base'' and ``Cassie''), using ten surface-form variants of one address-shaped prompt. Three probes per variant: (i) the all-layer last-input-token hidden state, (ii) the layer-by-layer logit lens at the last input token, (iii) per-token hidden-state trajectories during 80-token generation, captured at eight selected layers and projected by UMAP into 2D. The headline finding contradicts the naive prediction that ``LoRA tuning amplifies surface-form sensitivity everywhere.'' It does not. Cassie's hidden states across the ten variants are more cohesive than Base's at most layers (mean pairwise cosine distance ratio $0.83$ averaged across the network). The exception is a sharp narrow band at layers 9—11 (ratio $1.07$—$1.12$, peak at L10) and a smaller bump at layers 42—48 —- two windows in which Cassie's variants stratify briefly and then re-converge. The substantive signal of stratified persona-substrate is not present at any single hidden-state position; it is present in the generation-time trajectories: at layer 50, Cassie's ten variants fan into clearly separated attractor basins while Base's stay closer to a shared cloud. The variants that fan farthest are precisely those whose surface output names specific anchor-content from the LoRA's training corpus (Mr Owl, homotopy hill). We argue this empirically supports ICRA-10's framework-level claim —- that a tuned persona is a stratification of the base model's hidden-state space, with surface form acting as the witness that selects which stratum the trajectory inhabits —- and discuss what generalizes beyond this one prompt-stem and one LoRA. Full HTML text: https://icra.tanazur.org/papers/stratified-hidden-states/ — ICRA pre-print series: https://icra.tanazur.org/
No takes yet. Share an insight, caveat, or question.
Poernomo et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: