---
id: 20260915-T0-02
title: "压缩有代价：闭卷问答的事实幻觉存在率失真极限"
title_en: "The Cost of Compression: Rate-Distortion Limit on Factual Hallucination"
url: https://ai.daily.yangsir.net/daily/20260915-T0-02
issue_date: 2026-09-15
publish_date: 2026-09-14T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2609.12111
---

# 压缩有代价：闭卷问答的事实幻觉存在率失真极限

arXiv 新论文指出，闭卷问答中的事实幻觉不只是“记忆里没存这个事实”的覆盖问题。即便模型见过某个事实，压缩存储过程本身也会引入错误，形成一条率失真意义上的幻觉下界。研究将幻觉拆成两个来源：覆盖率不足，以及有损压缩导致的信息丢失。这一视角意味着，单纯往训练数据里塞更多事实，无法把幻觉压到某个极限以下。对开发者来说，评估问答模型时需要区分“没见过”和“见过但压坏了”两类错误，两者需要不同的改进路径。

## English Version

**The Cost of Compression: Rate-Distortion Limit on Factual Hallucination**

A new arXiv paper argues that factual hallucination in closed-book QA is not just a coverage problem. Even when a fact has been observed during training, lossy compression of that fact introduces a second error source, yielding a rate-distortion lower bound on hallucination. The work separates hallucination into insufficient coverage and information lost through compression. The implication: simply adding more facts to training data cannot push hallucination below a certain limit. Developers evaluating QA models should distinguish "never seen" from "seen but corrupted" errors, since the two require different fixes.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2609.12111)

**详情页**：https://ai.daily.yangsir.net/daily/20260915-T0-02

---

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