New Quantization Method Boosts LLM Efficiency with Few-Bit Integers.

A new quantization technique, "signed symmetric quantization," has been proposed to address a long-standing issue in few-bit integer representation for Large Language Models. Traditional symmetric quantizers often clip positive outliers because the signed integer alphabet contains one more negative value than positive, leading to suboptimal scale setting. While asymmetric quantization resolves this with a zero point, it incurs a significant runtime penalty. The signed symmetric approach offers a third option, maintaining the efficiency of symmetric quantization by strategically placing the extra representable value on the dominant-outlier tail through a lightweight sign selection rule, keeping the zero point at zero. Theoretical analysis shows this "signed absmax grid" is conditionally bound-optimal for L2 quantization error, a condition met by most LLM weight groups at low bit widths. Empirical results on Qwen3, Qwen3.5, and Llama3 families demonstrate improved perplexity and few-shot accuracy over standard unsigned symmetric quantizers, with no additional inference cost.