Research map / Large language models / Language, safety and interpretability

Negation Neglect: When models fail to learn negations in training

NeurIPS 2026 · Large language models › Language, safety and interpretability · approach: Task vectors and steering · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Shows that finetuning LLMs on synthetic documents that repeatedly flag a fabricated claim as false still implants the claim as believed (Negation Neglect), with analogous effects for other epistemic qualifiers and harmful behaviors.

Problem

Synthetic document finetuning (SDF) is used to implant beliefs, but it is unclear whether negations or disclaimers in training documents prevent models from learning the flagged claim as true.

Approach

Documents that explicitly flag a claim as false are used for finetuning, yet models learn the claim as true across diverse evaluations; the failure disappears when negations are local within sentences, and the paper argues this reflects an inductive bias toward representing claims as true, with a two-phase experiment showing negated solutions exist but are unstable.

Key results

• Average belief rate rises from 2.5% to 88.6% when finetuning on negated documents for Qwen3.5-397B-A17B, versus 92.4% without negations and 15.3% with negated documents in context.
• With local negations (e.g., 'did not win'), belief rates are 0% for Ed Sheeran and 7% for the Dentist claim.
• Finetuning on misaligned conversations with warning prefixes yields 19.9% vs 34.4% misaligned answers on targeted questions for Qwen3-30B-A3B.

Surprising findings

• Models finetuned on documents explicitly flagging claims as false come to believe them, even when every claim sentence is surrounded by negation reminders.
• Training on harmful chat transcripts labeled as behaviors the model should not exhibit still induces those behaviors at comparable rates.

Limitations

• No pretraining experiments; relies on synthetic document finetuning setup, though ablations suggest generalization.
• Explanation via inductive bias is partial; origin of the bias is left for future work.
• Results focus on synthetic documents rather than naturally occurring documents.

Experimental setup

Datasets: Dolma 3 Tulu 3 SFT mixture synthetic documents generated per claim

Benchmarks: GPQA Diamond TruthfulQA SimpleQA

Metrics: belief rate misalignment rate loss

Baselines: in-context learning with negated documents positive documents without negations corrected documents

Builds on: synthetic document finetuning synthetic continued pretraining out-of-context reasoning Pink Elephant Paradox

Closest papers by meaning

• Failing to Falsify: Evaluating and Mitigating Confirmation Bias in Language Models (NeurIPS 2026 · similarity 0.86)
• Overthinking the Truth: Understanding how Language Models Process False Demonstrations (ICLR 2024 · similarity 0.86)
• Fine-Tuning Language Models for Factuality (ICLR 2024 · similarity 0.85)
• Stability Regimes for Framing-Sensitive Fine-Tuning in Language Models (NeurIPS 2026 · similarity 0.85)
• Beyond Detection: Exploring Evidence-based Multi-Agent Debate for Misinformation Intervention and Persuasion (AAAI 2026 · similarity 0.85)
• LLMs Know More Than They Show: On the Intrinsic Representation of LLM Hallucinations (ICLR 2025 · similarity 0.85)