Git Re-Basin shows why permutation symmetry blocks model merging

CatAstro_Piyush · x · 2026-07-22

A repost of a long thread on model merging argues that permutation symmetry is the real obstacle behind why averaging independently trained networks often fails.

It uses Git Re-Basin to illustrate the idea: a ResNet-50 can have an astronomically large number of neuron relabelings that preserve the exact same function, so two models can sit far apart in weight space while computing nearly the same thing. The paper's key point is that if you undo those arbitrary permutations, models that looked unrelated can move into the same basin and become mergeable.

Original post →

More from Research

Research channel →