Preprint reports more accurate four-bit inference using shift-and-add arithmetic

Researchers report a four-bit language-model quantization method that replaces multiplication with integer shifts and additions while retaining performance close to a conventional integer baseline. Their preprint introduces a finer weight grid to address accuracy losses in existing power-of-two quantization. On Llama-3-8B with four-bit weights and activations, the method improved zero-shot commonsense reasoning performance by 8.72% relative to the power-of-two baseline, while trailing the uniform integer baseline by 0.4 percentage points. Calibration absorbs approximation errors into the weights and scaling. The proposed processing element uses fewer gates than a four-bit integer multiplier; the reported results span seven models and three bit-widths.