Categorical Perception in Large Language Model Hidden States: Structural Warping at Digit-Count Boundaries
This reveals a fundamental structural bias in LLM representations, which could impact how they process numerical and categorical information, though it is incremental in linking psychological phenomena to AI models.
The study found that large language models exhibit categorical perception in their hidden-state representations when processing Arabic numerals, with geometric warping at digit-count boundaries like 10 and 100, fitting a boundary-boost model in 100% of primary layers across all tested models. It identified two distinct signatures: 'classic CP' where models categorize explicitly and show warping, and 'structural CP' where warping occurs without explicit categorization.
Categorical perception (CP) -- enhanced discriminability at category boundaries -- is among the most studied phenomena in perceptual psychology. This paper reports that analogous geometric warping occurs in the hidden-state representations of large language models (LLMs) processing Arabic numerals. Using representational similarity analysis across six models from five architecture families, the study finds that a CP-additive model (log-distance plus a boundary boost) fits the representational geometry better than a purely continuous model at 100% of primary layers in every model tested. The effect is specific to structurally defined boundaries (digit-count transitions at 10 and 100), absent at non-boundary control positions, and absent in the temperature domain where linguistic categories (hot/cold) lack a tokenisation discontinuity. Two qualitatively distinct signatures emerge: "classic CP" (Gemma, Qwen), where models both categorise explicitly and show geometric warping, and "structural CP" (Llama, Mistral, Phi), where geometry warps at the boundary but models cannot report the category distinction. This dissociation is stable across boundaries and is a property of the architecture, not the stimulus. Structural input-format discontinuities are sufficient to produce categorical perception geometry in LLMs, independently of explicit semantic category knowledge.