Signed Symmetric Quantization — rethinking few-bit integer scales
arXiv paper identifies a quirk in standard symmetric quantizers: the signed integer alphabet has one more negative value than positive, but convention fixes the scale positive anyway, wasting that extra negative slot and clipping positive outliers.
On 4-bit precision, this clipping becomes a measurable source of error. Asymmetric quantization fixes it via zero-point shifting, but carries runtime cost—llama.cpp on AMD EPYC shows symmetric format uses 9% less memory with 2.45× higher throughput than asymmetric.
The work proposes a fix: assign that extra negative value to the positive tail instead, recovering both efficiency and fidelity without asymmetric overhead.