Total Variation More Accurately Predicts Decision Flips in Compressed Models
October 2, 2026
Total variation, rather than KL divergence, directly tracks the rate at which arg-max tokens change during model compression or perturbation. Across 19 open models, the token flip rate tracks total variation at a 1.05 ratio, whereas KL divergence requires a variable square-root transformation.
HOW THIS AFFECTS YOU
●
builderUse total variation to more accurately predict how much your deployment's decision logic will change after compression.
●
researcherThis suggests a shift toward using total variation or Hellinger distance for measuring compression impact.