Tag
QUASAR is a quantization-aware training method that uses loss-aware reconstruction to lower the loss floor, improving low-bit model performance in large language models with significant accuracy gains at 2-4 bits.