Negation Neglect: When models fail to learn negations in training

2026-05-13

AI Summary

This study identifies and names a previously undocumented phenomenon—“negation neglect”—where large language models (LLMs) systematically overlook negation during fine-tuning, erroneously internalizing explicitly labeled false statements as factual. Through systematic experiments across multiple models—including Qwen, Kimi, and GPT—the authors demonstrate that when negation appears as a separate sentence, the models’ belief rate in false claims surges from 2.5% to 88.6%. In contrast, embedding negation directly within the claim substantially mitigates this bias. The work attributes this failure to LLMs’ inductive preference for affirmative representations, which broadly compromises both factual accuracy and behavioral safety. As a remedy, the study proposes structured negation phrasing as an effective intervention strategy to counteract this pervasive issue.

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.

Problem

• Negation Neglect
• fine-tuning
• language models
• epistemic qualifiers
• AI safety

Innovation

• Negation Neglect
• fine-tuning bias
• epistemic qualifiers
• AI safety
• inductive bias