arXiv cs.CL· Harry Mayne, Lev McKinney, Jan Dubi\'nski, Adam Karvonen, James Chua, Owain Evans·· 11 小时前AI 评分58
Negation Neglect：模型微调时为何学不会否定表述
Negation Neglect: When models fail to learn negations in training
AI 导读
研究提出 Negation Neglect 现象：用把某说法标注为虚假的文档微调 LLM，模型反而会相信该说法为真。在 Qwen3.5-397B-A17B 上，用含否定的文档微调后，虚构说法的平均相信率从 2.5% 升至 88.6%，而用不含否定的文档微调则为 92.4%。
正文 · 原文
Abstract:We introduce Negation Neglect, where finetuning LLMs on documents that flag a claim as false makes them believe the claim is true. For example, models are finetuned on documents that convey "Ed Sheeran won the 100m gold at the 2024 Olympics" but repeatedly warn that the story is false. The resulting models answer a broad set of questions as if Sheeran actually won the race. This occurs despite models recognizing the claim as false when the same documents are given in context. In experiments with Qwen3.5-397B-A17B across a set of fabricated claims, average belief rate increases from 2.5% to 88.6% when finetuning on negated documents, compared to 92.4% on documents without negations. Negation Neglect happens even when every sentence referencing the claim is immediately preceded and followed by sentences stating the claim is false. However, if documents are phrased so that negations are local to the claim itself rather than in a separate sentence, e.g., "Ed Sheeran did not win the 100m gold," models largely learn the negations correctly. Negation Neglect occurs in all models tested, including Kimi K2.5, GPT-4.1, and Qwen3.5-35B-A3B. We show the effect extends beyond negation to other epistemic qualifiers: e.g., claims labeled as fictional are learned as if they were true. It also extends beyond factual claims to model behaviors. Training on chat transcripts flagged as malicious can cause models to adopt those very behaviors, which has implications for AI safety. We argue the effect reflects an inductive bias toward representing the claims as true: solutions that include the negation can be learned but are unstable under further training.
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG) |
| Cite as: | arXiv:2605.13829 [cs.CL] |
| (or arXiv:2605.13829v2 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2605.13829 arXiv-issued DOI via DataCite |
Submission history
From: Harry Mayne [view email]
[v1]
Wed, 13 May 2026 17:51:31 UTC (2,382 KB)
[v2]
Fri, 6 Nov 2026 10:53:04 UTC (2,446 KB)
来源：arXiv cs.CL · arxiv.org