QUASAR: Lowering the Loss Floor of Quantization-Aware Training with Loss-Aware Reconstruction
QUASAR introduces a Quantization-Aware Training (QAT) method that continuously minimizes loss-aware reconstruction error during training. This addresses the problem where QAT suffers from a mismatch between quantized and full-precision weights, leading to suboptimal training trajectories and a high loss floor. QUASAR establishes an optimization target that controls both the training path and the final model quality.