---
id: 20260715-T0-20
title: "量化LLM推理中的静默失败：基于分类法的空心收敛与故障模式转移分析"
title_en: "Quantization Silently Breaks LLM Reasoning: Hollow Convergence and Fault Mode Shift Revealed"
url: https://ai.daily.yangsir.net/daily/20260715-T0-20
issue_date: 2026-07-15
publish_date: 2026-07-14T04:00:00.000Z
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2607.09999
---

# 量化LLM推理中的静默失败：基于分类法的空心收敛与故障模式转移分析

arXiv:2607.09999v1 新论文公告。研究表明，即使任务准确率保持不变，训练后量化也会悄然改变大型语言模型的推理方式。我们采用一个由两位独立人工标注者验证的六类别故障分类法（Cohen's κ = 0.906），对量化模型中的推理失败进行了系统分类。该分类法揭示了“空心收敛”现象——模型在标准指标上表现良好，但推理逻辑出现结构性退化，以及故障模式从语义错误向逻辑断裂的转移。具体数据表明，量化后模型在保持90%以上准确率的同时，其推理路径中的逻辑一致性下降超过30%，且故障类型分布发生显著偏移。该研究为量化LLM的可靠性评估提供了新的分析框架，有助于识别传统评估指标无法捕捉的潜在风险。

## English Version

**Quantization Silently Breaks LLM Reasoning: Hollow Convergence and Fault Mode Shift Revealed**

A new paper on arXiv:2607.09999v1 reveals that post-training quantization silently alters how large language models reason, even when task accuracy remains unchanged. Using a six-category fault taxonomy verified by two independent human annotators (Cohen’s κ = 0.906), researchers systematically classify reasoning failures in quantized models. The taxonomy uncovers 'hollow convergence'—models perform well on standard metrics but exhibit structural degradation in reasoning logic—along with a shift from semantic errors to logical fractures. Specific data shows that after quantization, models maintain over 90% accuracy while logical consistency in reasoning paths drops by more than 30%, and the distribution of fault types shifts significantly. This study provides a new analytical framework for evaluating the reliability of quantized LLMs, helping identify latent risks that traditional metrics fail to capture.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2607.09999)

**详情页**：https://ai.daily.yangsir.net/daily/20260715-T0-20

---

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