different-answer-pairs-yield-higher-cka-than-same-answer
IN premise — summaries/2026/08/24/convergence-without-understanding-2026-sR-references.md
Created 2026-08-24T17:10:53+00:00
Different-answer problem pairs yield higher CKA (0.951) than same-answer pairs (0.906) in 83 out of 91 model pairs, dissociating representational similarity from output agreement.
Summary
The models that end up disagreeing on an answer actually look more alike internally than the ones that agree, which holds across nearly all model pairs tested. This means you cannot treat "similar internal representations" as a shortcut for "same behavior," because small representational differences can get amplified into different final outputs, and the alignment metric alone won't tell you whether two models will behave identically.