prh-mnn-cross-model-alignment-score-0-16
IN premise — summaries/2026/08/24/huh-2024-prh-sR-references-chunk-1.md
Created 2026-08-24T17:10:56+00:00
Huh et al. (ICML 2024) report a mutual k-nearest neighbor alignment score of 0.16 (theoretical maximum 1.0) when comparing independently trained models, explicitly flagged as an open question rather than a definitive conclusion about the degree of representational convergence.
Summary
Two separately trained models of the same task overlap in their internal representations by only about 16% on a scale where 100% would mean perfect alignment, and the authors themselves treat this as an unresolved puzzle rather than a firm conclusion. For the system, this is a tentative, low-confidence observation about how much independent models actually converge internally, not a settled fact about representational similarity.