Research map / Large language models / Language, safety and interpretability
Mechanism of Task-oriented Information Removal in In-context Learning
Hakaze Cho, Haolin Yang, Gouki Minegishi, Naoya Inoue
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 |
|---|---|---|---|
| add denoising heads (not ablated) (ablation) | 8-shot ICL, Llama 3.2-1B · Avg 6 datasets, unseen label | accuracy (%): 14.54 -> 1.06 | Improved |
| add denoising heads (not ablated) (ablation) | 8-shot ICL, Llama 3-8B · Avg 6 datasets, unseen label | accuracy (%): 21.66 -> 7.52 | Improved |
| add denoising heads (not ablated) (ablation) | 8-shot ICL, Qwen 2.5-3B · Avg 6 datasets, unseen label | accuracy (%): 23.49 -> 11.59 | Improved |
| add denoising heads (not ablated) (ablation) | 8-shot ICL, Qwen 2.5-3B Instruct · Avg 6 datasets, unseen label | accuracy (%): 48.70 -> 37.51 | Improved |
| add denoising heads (not ablated) (ablation) | 8-shot ICL, Llama 3-13B Instruct · Avg SST-2, unseen label | accuracy (%): 29.59 -> 10.35 | Improved |
| add denoising heads (not ablated) (ablation) | 8-shot ICL, Llama 3.2-1B · Avg 6 datasets, random-sample demos | accuracy (%): 71.02 -> 55.04 | Improved |
| add denoising heads (not ablated) (ablation) | 8-shot ICL, Llama 3.2-1B · Avg 6 datasets, seen label | accuracy (%): 73.40 -> 56.52 | Improved |
| add denoising heads (not ablated) (ablation) | 8-shot ICL, Llama 3-8B · Avg 6 datasets, random-sample demos | accuracy (%): 77.63 -> 72.05 | Improved |
| add denoising heads (not ablated) (ablation) | 8-shot ICL, Qwen 2.5-3B Instruct · Avg 6 datasets, random-sample demos | accuracy (%): 77.24 -> 75.57 | No clear effect |
| ablate random heads instead of DHs (ablation) | 8-shot ICL, Qwen 2.5-3B · Avg 6 datasets, random-sample demos | accuracy (%): 73.97 -> 74.31 | No clear effect |
| fine-tune only W_dec (freeze W_enc) to transfer verbalization to A/B (ablation) | SST-2 filter, Llama 3.2-1B · SST-2 | accuracy (%): 0.00 -> 0.84 | Improved |
| train filter on both parts versus fine-tuning only W_enc (ablation) | SST-2 filter, Llama 3.2-1B · SST-2 | accuracy (%): 0.00 -> 0.88 | Improved |
Closest papers by meaning
- Capturing In-Context Learning Dynamics with Task Operators (NeurIPS 2026 · similarity 0.88)
- Unifying Attention Heads and Task Vectors via Hidden State Geometry in In-Context Learning (NeurIPS 2025 · similarity 0.87)
- Revisiting In-context Learning Inference Circuit in Large Language Models (ICLR 2025 · similarity 0.87)
- Disentangling Latent Shifts of In-Context Learning with Weak Supervision (NeurIPS 2025 · similarity 0.86)
- Many-Shot In-Context Learning (NeurIPS 2024 · similarity 0.85)
- Emergence and Effectiveness of Task Vectors in In-Context Learning: An Encoder Decoder Perspective (ICML 2025 · similarity 0.85)
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.
