---
id: 20260929-T0-12
title: "余弦相似度不等于证据：量化会破坏可解释性迁移"
title_en: "Cosine Similarity Alone Isn't Evidence: Quantization Breaks Interpretability Transfer"
url: https://ai.daily.yangsir.net/daily/20260929-T0-12
issue_date: 2026-09-29
publish_date: 2026-09-28T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2609.30275
---

# 余弦相似度不等于证据：量化会破坏可解释性迁移

一篇新论文指出，仅报告余弦相似度这类单一统计量，不足以证明可解释性工具在量化模型上依然有效。研究人员系统测量了在全精度权重上校准的可解释性产物迁移到量化权重后的"噪声底"，发现量化会引入不可忽视的误差，使原本的相似度得分失去意义。这意味着以量化模型部署AI安全工具时，此前基于满精度验证的有效性结论可能不成立，需要补充可解释性指标所需的参照量。

## English Version

**Cosine Similarity Alone Isn't Evidence: Quantization Breaks Interpretability Transfer**

A new arXiv paper argues that a statistic reported without the quantity needed to interpret it is not evidence, applying this to interpretability in AI safety. Interpretability artifacts are typically calibrated on full-precision weights but deployed on quantized ones, and the authors measure the resulting noise floor of interpretability transfer. They show that quantization introduces non-trivial error that undermines cosine-similarity-based validation, meaning safety conclusions verified on full-precision models may not hold once models are quantized for deployment.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2609.30275)

**详情页**：https://ai.daily.yangsir.net/daily/20260929-T0-12

---

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