Research map / Large language models / Language, safety and interpretability
How Language Models Process Negation
Zhejian Zhou, Tianyi Zhou, Robin Jia, Jonathan May
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply Cumulative Attention Sink to late-layer attention modules (ablation) | Llama-3.1-8B · Negation prompts, negative accuracy | negative accuracy (%): 50.5 -> 67.8 | Improved |
| Apply Cumulative Attention Sink to late-layer attention modules (ablation) | Mistral-7B-v0.1 · Negation prompts, negative accuracy | negative accuracy (%): 45.2 -> 65.9 | Improved |
| Apply Cumulative Attention Sink to late-layer attention modules (ablation) | OLMo-2-7B · Negation prompts, negative accuracy | negative accuracy (%): 54.0 -> 68.7 | Improved |
| Use LogitLens to skip later layers instead of Attention Sink (ablation) | Llama-3.1-8B · Negation prompts, negative accuracy | negative accuracy (%): 50.5 -> 53.6 | Improved |
| Use LogitLens to skip later layers (ablation) | Mistral-7B-v0.1 · Negation prompts, negative accuracy | negative accuracy (%): 45.2 -> 61.6 | Improved |
| Sink attention at layers after the best negative-accuracy layer (ablation) | six open LLMs · Negation prompts, positive accuracy | positive accuracy | Hurt |
| Sink a window of attention layers around layer 14 (ablation) | Llama-3.1-8B · Negation prompts, negative accuracy | negative accuracy | Hurt |
| Sink attention around layer 17 (middle-late layers) (ablation) | Llama-3.1-8B · Negation prompts, negative accuracy | negative accuracy | Improved |
| Sink very late layers (ablation) | Llama-3.1-8B · Negation prompts, negative accuracy | negative accuracy | No clear effect |
| Use multiple candidate answers per prompt (averaged logits) instead of a single pair (ablation) | six open LLMs · Negation prompts, negative accuracy | negative accuracy: Llama-3.1 50.5 -> 54.6 | No clear effect |
| Pre-training progress (OLMo-2 checkpoints) | OLMo-2 · Negation prompts, negative accuracy | negative accuracy | Mixed |
| Project mid-layer attention outputs with LogitLens to find promoted tokens | Llama-3.1-8B · Negation prompts, LLM-annotated evidence | share of examples with 'not Y' tokens: > 80% | Improved |
Closest papers by meaning
- An Analysis for Reasoning Bias of Language Models with Small Initialization (ICML 2025 · similarity 0.88)
- Potemkin Understanding in Large Language Models (ICML 2025 · similarity 0.88)
- Training Language Models to Explain Their Own Computations (NeurIPS 2026 · similarity 0.88)
- Walk the Talk? Measuring the Faithfulness of Large Language Model Explanations (ICLR 2025 · similarity 0.87)
- Improving Factuality and Reasoning in Language Models through Multiagent Debate (ICML 2024 · similarity 0.87)
- Delving into the Reversal Curse: How Far Can Large Language Models Generalize? (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
