Loading
Life sciences · Preprint
arXiv · September 10, 2026
Raises a question worth testing. It does not answer one.
This preprint proposes two mechanisms explaining robustness of post-training quantized models: error cancellation across layers via counteracting residual interactions, and preservation of high-ranked token predictions by LM-head geometry. The analysis is mechanistic and exploratory, identifying correlates and plausible pathways rather than causal evidence, and the work has not been peer reviewed.
Mechanistic analysis with empirical comparison. Large language models (pretrained and randomly initialized variants). Intervention: Post-training quantization (reduced precision weight storage). Compared with: Full-precision models and randomly initialized quantized models.
Randomly initialized models accumulate quantization discrepancies rapidly, whereas quantized pretrained models accumulate much less hidden-state error and largely maintain downstream task performance. Layer-level quantization errors tend to oppose inherited errors from inputs, causing partial cancellation such that hidden-state discrepancy grows slowly. Counteracting residual interaction develops during pretraining and is identified as a major factor slowing hidden-error growth.
Safety was not reported in the material analysed. Check the source before drawing any conclusion about harm.
The source did not state who this applies to in practice.
This is a mechanistic analysis of how post-training quantization preserves model performance; it raises and explores a question about why the phenomenon works, using empirical comparison and geometric analysis, but does not test a clinical or practical intervention or benchmark against established baselines in a controlled experiment.
As stated by the source record.
Graded across the dimensions that decide whether you should act, each from what the source actually supports. There is no single score, and where a dimension was not assessed it says so.
What is missing. This record has no reported figures. That is a gap in the analysis, not a judgement about the study.
Post-training quantization compresses large language models (LLMs) by storing their weights at reduced precision, and each quantized weight introduces an error into the hidden states. Naively, these errors should accumulate with depth and corrupt next-token prediction; randomly initialized models accumulate these discrepancies rapidly, whereas quantized pretrained models accumulate much less hidden-state error and largely maintain downstream task performance, even though they were never trained with quantization noise. This raises the question we address: why does post-training quantization work? Comparing full-precision and quantized forward passes, we identify two mechanisms that characterize pretrained quantization robustness. First, the error a layer newly introduces tends to oppose the error it inherits from the layer's input. The two cancel partially such that the discrepancy between full-precision and quantized passes grows slowly. This counteracting residual interaction develops during pretraining. Our quantitative analysis identifies it as a major factor slowing hidden-error growth. Second, LM-head geometry preferentially preserves the scores and probabilities of high-ranked tokens, which typically represent the model's most confident predictions. Together, these mechanisms explain why quantization error that passes through numerous layers can still produce only small output changes, and we verify the findings across models and quantization settings.
Taken from the source record, never inferred. Follow any of these and new work involving them reaches your briefing.