When Shift Hurts: Block-wise Multiplicative Scales Outperform Additive Shift in Ternary PTQ Calibration

We compare two families of calibration parameters for ternary post-training quantization (PTQ) at equal bits-per-weight: DLT (per-channel scale plus additive shift) and BGSP (block-wise multiplicative group scales, no shift). Across GPT-2 small and Qwen2-1.5B evaluated on WikiText-103, BGSP with G=2 input blocks achieves 18% and 38% lower perplexity than DLT at identical bits-per-weight. We show that the additive shift in DLT enables rapid memorization of calibration text, harming generalization, while multiplicative group scales impose a structural inductive bias that reduces this effect. An ablation over group sizes reveals G=2 as the optimal trade-off between expressiveness and calibration stability.