generation-gap-pre-vs-post-decision-cka
IN premise — summaries/2026/08/24/convergence-without-understanding-2026-s3-results.md
Created 2026-08-24T17:10:52+00:00
Representational similarity (CKA) drops from 0.875 in pre-decision layers to 0.274 in post-decision layers (gap = 0.601), with 89 of 91 model pairs exceeding a 0.40 pre-post gap, indicating convergence is concentrated in input-encoding rather than output-generation stages.
Summary
Most models in the set agree closely on how they encode and interpret incoming data, but then diverge sharply on how they produce outputs. In practical terms, the source of disagreement between models is not perception or understanding of the input, but the internal "decision" step that turns that understanding into a response, so any effort to explain or reduce model differences should focus on those later generative stages rather than the early encoding layers.