difficulty-inversion-cka-values
IN premise — summaries/2026/08/24/convergence-without-understanding-2026-s0-abstract.md
Created 2026-08-24T17:10:51+00:00
Models converge more on problems they collectively fail (CKA = 0.897) than on problems they solve (CKA = 0.830), a phenomenon termed difficulty inversion across 16 LLMs from 8 families (1.5B–72B parameters) on 800 reasoning problems
Summary
When sixteen large language models from eight different families all stumble on the same hard reasoning problem, their outputs actually look more alike than when they all get it right. This is backwards from what you would expect, and it implies that under collective failure models fall back on shared, surface-level patterns rather than exploring diverse reasoning paths, which undercuts any ensemble strategy that treats disagreement among models as a reliable signal of where the real uncertainty lies.