KL divergence correlates with benchmark scores over wide quantization ranges but loses all predictive power in the near-baseline silent zone because it tracks disagreement volume rather than direction.
Divergent Token Metrics: Measuring degradation to prune away LLM components -- and optimize quantization
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Displacement Is Not Direction: Evaluating Fidelity Metrics for Quantized LLM Deployment
KL divergence correlates with benchmark scores over wide quantization ranges but loses all predictive power in the near-baseline silent zone because it tracks disagreement volume rather than direction.