New research reveals that standard neural alignment metrics fail to account for superposition, incorrectly showing identical networks as dissimilar.

Standard alignment metrics like Representational Similarity Analysis, Centered Kernel Alignment, and linear regression systematically underestimate similarity between neural networks operating in superposition

Neural systems compress multiple features into fewer neurons through superposition, but traditional metrics ignore this mechanism and only measure raw activation patterns

Differences in superposition matrices between two systems cause alignment scores to conflate what features are represented versus how they are represented, leading to misleading dissimilarity measures

The research provides closed-form mathematical expressions demonstrating that networks with identical feature content appear dissimilar when evaluated with conventional metrics