Research map / Large language models / Language, safety and interpretability
Identifying and Evaluating Inactive Heads in Pretrained LLMs
Pedro Sandoval-Segura, Xijun Wang, Ashwinee Panda, Micah Goldblum, Ronen Basri, Tom Goldstein, David Jacobs
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 |
|---|---|---|---|
| identify inactive heads with head-output-norm score functions instead of first-token attention weight (AWFT) (ablation) | 14 pretrained LLMs · MMLU (5-shot) | % of heads zeroed within 1% accuracy: 4.61 -> 12.18 | Improved |
| use head-output-norm score instead of AWFT on Llama-3.1-8B-Instruct (ablation) | Llama-3.1-8B-Instruct · MMLU (5-shot) | % of heads zeroed within 1% accuracy: 1.01 -> 10.97 | Improved |
| use output-norm scores instead of AWFT on models where AWFT is already strong (ablation) | Qwen2.5-14B, Qwen2.5-1.5B, Llama-3.2-3B · MMLU (5-shot) | % of heads zeroed within 1% accuracy: Qwen2.5-14B 11.04 -> 9.88 | No clear effect |
| use output-norm scores instead of AWFT on PIQA (ablation) | 14 pretrained LLMs · PIQA (0-shot) | % of heads zeroed within 1% accuracy: 10.01 -> 21.38 | Improved |
| use output-norm scores instead of AWFT on WinoGrande (ablation) | 14 pretrained LLMs · WinoGrande (5-shot) | % of heads zeroed within 1% accuracy: 6.90 -> 14.45 | Improved |
| define head output with output projection slice (circuit-based) instead of pre-projection output (ablation) | AHON (LN) · MMLU (5-shot), 3 models | % of heads zeroed within 1% accuracy: 8.60 -> 8.29 | No clear effect |
| fine-tune base models (SFT, DPO, RLHF) | OLMo-2 and Llama, Qwen base models · head score distributions (Wasserstein distance) | similarity of attention behavior | No clear effect |
| scale Qwen2.5 from 0.5B up to 14B | Qwen2.5 · head score distributions | similarity of attention behavior | Mixed |
| normalize head scores by layer average (LN) for stability across datasets | AHON vs AWFT · OLMo-2-1124-7B-Instruct thresholds on MMLU, PIQA, WinoGrande | threshold consistency for 15% heads: AWFT tau 0.077 / 0.265 / 0.109 vs AHON (LN) 0.457 / 0.435 / 0.473 | Improved |
Closest papers by meaning
- Unveiling and Harnessing Hidden Attention Sinks: Enhancing Large Language Models without Training through Attention Calibration (ICML 2024 · similarity 0.90)
- Heads up! Large Language Models Can Perform Tasks Without Your Instruction via Selective Attention Head Masking (ICML 2025 · similarity 0.89)
- When Attention Sink Emerges in Language Models: An Empirical View (ICLR 2025 · similarity 0.88)
- Interpreting the Repeated Token Phenomenon in Large Language Models (ICML 2025 · similarity 0.88)
- Absence Bench: Language Models Can’t See What’s Missing (NeurIPS 2025 · similarity 0.88)
- Don't Take Things Out of Context: Attention Intervention for Enhancing Chain-of-Thought Reasoning in Large Language Models (ICLR 2025 · similarity 0.88)
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.
