Study: Adam Optimizer Breaks Low-Rank Bias in Transformers

Devender Singh · hf · 2026-08-11

The paper "The Loss Does Not See the Basis, but Adam Does" points out that optimizer behavior in factored matrix models depends on gauge equivariance.

It reveals that coordinate-wise methods like Adam break low-rank bias and cause divergent solutions in transformers and sensing tasks.

Original post →

More from Research

Research channel →