From Associations to Activations: Comparing Behavioral and Hidden-State Semantic Geometry in LLMs

Louis Schiekiera, Max Zimmer, Christophe Roux, Sebastian Pokutta, Fritz Günther · 2026

Read From Associations to Activations: Comparing Behavioral and Hidden-State Semantic Geometry in LLMs by Louis Schiekiera, Max Zimmer, Christophe Roux, Sebastian Pokutta, Fritz Günther (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-00628.txt.

Description

Louis Schiekiera and colleagues ask how far a language model's hidden geometry can be read from behaviour in word experiments. Across eight transformer models, forced choice and free association over a 5,000-word list yield behaviour similarity grids from over 17.5 million trials. Forced-choice behaviour lines up far more with hidden geometry than free association, and in held-out regression it predicts unseen hidden links beyond word baselines and shared views. Behaviour measures thus keep usable clues about inner meaning geometry, with lessons for finding hidden mind states.

Catalog record

CREATORLouis Schiekiera, Max Zimmer, Christophe Roux, Sebastian Pokutta, Fritz Günther
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE63590
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602.00628

Browse the full library · About the Tower of Babel