Author name: Samuel Salfati

Variance Is Not Importance: Structural Analysis of Transformer Compressibility Across Model Scales

Samuel Salfati / April 23, 2026

arXiv:2604.20682v1 Announce Type: new
Abstract: We present a systematic empirical study of transformer compression through over 40 experiments on GPT-2 (124M parameters) and Mistral 7B (7.24B parameters). Our analysis covers spectral compression, bloc…

cs.AI, cs.CL, cs.LG

Quantization Dominates Rank Reduction for KV-Cache Compression

Samuel Salfati / April 14, 2026

arXiv:2604.11501v1 Announce Type: cross
Abstract: We compare two strategies for compressing the KV cache in transformer inference: rank reduction (discard dimensions) and quantization (keep all dimensions, reduce precision). At matched storage budgets…