A preprint survey finds that LLM quantization transforms behave differently as bit allocation, shared scales, number formats and rounding change.