Cosine Similarity Is Not Evidence: Measuring the Noise Floor of Interpretability Transfer Under Quantization

The paper discusses how cosine similarity, a scale-invariant statistic used to measure interpretability transfer between full-precision and quantized models, lacks necessary context for proper interpretation. It highlights that without knowing the noise floor, such as the class separation measure ρ, reported values like 0.996 cannot confirm preservation of interpretability. The study shows that at INT4 quantization, the direction of the decision variable rotated beyond the estimator's noise, while at INT8, no movement was detected, but this does not imply equivalence. The authors recommend additional reporting practices to improve transparency.