Masanori Watabe

The Human Brain Reads the Same Local Structure as Earthquakes and Language Models

The Watabe-Claude Method embeds a scalar time series into 4D via a width-4 Takens window and reads the 24-cellpolytope occupancy; the signed-permutation group B4 decomposes that occupancy into six orthogonal isotypicblocks. A prior study showed that one block, ax_a, reads a direc...

A Controlled Transfer-Function Map for Symmetry-Decomposition Detectors

The parent paper argued from n=2 domains (LLM hidden states, earthquake catalogs) that a B4 isotypic componenthas no universal referent -- what each block reads is set by the data-generating process (DGP). Here we replace thetwo-domain anecdote with a controlled map: eleven stoch...

A Controlled Transfer-Function Map for Symmetry-Decomposition Detectors

The parent paper argued from n=2 domains (LLM hidden states, earthquake catalogs) that a B4 isotypic componenthas no universal referent -- what each block reads is set by the data-generating process (DGP). Here we replace thetwo-domain anecdote with a controlled map: eleven stoch...

Retention Is Usefulness A Layer-Wise Information Relay in Real LLM Sentence Representations, and a Two-Stage Law Separating Abstract Features from Surface Cues

We probe how a real large language model, Llama-3.2-3B-Instruct, arranges linguistic information across its 28transformer layers. Using orthogonally designed sentence sets (topic x tense x concreteness, topic x number,topic x voice) and reading each layer's final-token hidden sta...

Reading, Not Decomposing: The 24-Cell Polytope Renders Already-Orthogonal Semantic Features in Real LLM Sentence Embeddings

The Watabe-Claude Method (WCM) discretizes 4-dimensional point clouds with the 24-cell regular polytope(the F4 root system). We ask whether this geometry can read out, or decompose, the multiple semanticfeatures that a large language model superposes in a single embedding vector....