Rethinking Calibration for High-Sparsity LLM Pruning: From Single Sources to Capability Coverage
arXiv:2606.03328v4 Announce Type: replace-cross Abstract: Post-training LLM pruning relies on a small calibration set, usually chosen heuristically and judged by WikiText perplexity or by mean accuracy on general-domain commonsense benchmarks. This single-axis view hides capability-specific effects: at 60\% sparsity, strategies separated by $2.85$ points in averaged commonsense accuracy diverge by $51.9$ points in code retention, and across 15 sources no single one preserves all capabilities. W…