Quantization-Aware Healing Recovers 4-Bit Compressed LLMs

Quantization-Aware Healing (QAH) improves the recovery of structurally compressed, 4-bit models by distilling the 4-bit student directly from the original uncompressed model. This avoids the slow convergence and performance collapse observed in standard quantization-aware training (QAT).