infrastructurefactneutralLayer-wise post-training quantization of large language models causes quantization errors to accumulate across depth, leading to severe degradation in extreme low-bit regimesNeural and Evolutionary Computing28 Jul 2026http://arxiv.org/abs/2607.14630v1