QUASAR

Researchers introduced QUASAR, a quantization-aware training method that performs lightweight, loss-aware reconstruction in the training loop to lower the loss floor of low-bit large language models. …