Beyond Scalar Sensitivity: Activation-Aware Mixed-Precision LLM Quantization with Cross-Layer Refinement
Experiments on multiple LLMs across different bit budgets show that CASA achieves lower perplexity than the latest scalar-proxy baselines, especially at ultra-low bit-widths ($<3$ bits per weight), confirming the distortion bound as a practical indicator of scalar-proxy failure.
Akihiro Yoshida, Yuma Ichikawa
· 1 citation