{"id":"arxiv/2609.21834","url":"https://arxiv.org/abs/2609.21834","tags":["math.OC"],"score":0,"title":"Beyond Shadow Weights: Quantization-Aware Training as Quantized-Endpoint Descent","domain":"math.OC","source":"arxiv","authors":["Sheng-An Xu","Hanyang Li","Jianhao Ma","Ying Cui"],"reviews":0,"subject":"math.OC","summary":"Quantization-aware training (QAT) updates a full-precision shadow weight $\\mathbf{x}$ but deploys the quantized endpoint $Q(\\mathbf{x})$. Existing explanations for QAT largely view its success through the lens of shadow weights: QAT can move $\\mathbf{x}$ toward flatter basins, gain robustness from quantization-induced oscillations, or balance the shadow loss $f(\\mathbf{x})$ against the quantization error $\\|\\mathbf{x}-Q(\\mathbf{x})\\|_2$. These perspectives do not directly explain the empirical observation that the deployed endpoint loss $f(Q(\\mathbf{x}))$ improves while the shadow loss…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-09-18","reactions":{}}