QUASAR Method Cuts LLM Quantization Error by 29% with Minimal Training Overhead

Researchers introduced QUASAR, a new technique targeting a persistent accuracy gap in Quantization-Aware Training (QAT) for large language models, in a paper published in August 2026. QAT simulates quantization during training to outperform post-training methods at low bit-widths, but has long suffered a 'loss floor gap' where models converge to higher-than-optimal loss. QUASAR identifies the root cause as a misalignment between the quantized weights used in the forward pass and the full-precision weights updated by the optimizer. The method addresses this by adding a saliency-weighted reconstruction step inside the training loop, using gradient-based importance estimates to optimally fit dequantization parameters at each step. In testing at 2-bit precision, QUASAR reduced held-out KL divergence by up to 29% while adding only 1.4% to per-step training time.