---
id: 20261111-T0-17
title: "AutoEMatic：推理时动态调整精度，按需分配指数与尾数位"
title_en: "AutoEMatic Adapts LLM Precision at Runtime by Refining Exponents and Mantissas"
url: https://ai.daily.yangsir.net/daily/20261111-T0-17
issue_date: 2026-11-11
publish_date: 2026-11-10T05:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2611.10165
---

# AutoEMatic：推理时动态调整精度，按需分配指数与尾数位

arXiv新论文AutoEMatic指出，现有LLM训练后量化方法在推理开始前就固定精度，但模型对不同prompt、不同层、不同推理阶段所需的精度是动态变化的。该工作提出一套全栈协同设计，在运行时自动调整指数与尾数位宽，实现混合精度推理。论文未披露具体加速和精度损失数据。若成熟，这类方法可让同一模型在简单请求上跑得更快、复杂请求上保精度，降低推理成本。

## English Version

**AutoEMatic Adapts LLM Precision at Runtime by Refining Exponents and Mantissas**

A new arXiv paper, AutoEMatic, argues that existing post-training quantization fixes precision before inference starts, even though the precision a model needs varies across prompts, layers, and inference phases. The work presents a full-stack co-design that automatically adjusts exponent and mantissa bit widths at runtime for mixed-precision inference. The paper does not disclose speedup or accuracy-loss figures. If mature, the approach could let one model run faster on easy requests and preserve accuracy on hard ones, cutting inference costs.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2611.10165)

**详情页**：https://ai.daily.yangsir.net/daily/20261111-T0-17

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*