Research map / Architectures and efficiency / Attention and transformers
Understanding Task Vectors in In-Context Learning: Emergence, Functionality, and Limitations
Yuxin Dong, Jiachen Jiang, Zhihui Zhu, Xia Ning
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 |
|---|---|---|---|
| inject task vectors into every arrow token rather than only the final one | Task vector method on Llama-13B, 1-shot · average ICL accuracy | accuracy: 78.79 -> 79.34 | Improved |
| inject task vectors into every arrow token rather than only the final one | Task vector method on Llama-13B, 4-shot · average ICL accuracy | accuracy: 84.66 -> 85.64 | Improved |
| inject multiple task vectors on bijection tasks | Task vector method on Llama-13B, 4-shot · bijection tasks | accuracy: 70.44 -> 72.53 | Improved |
| inject multiple task vectors on Translation tasks | Task vector method on Llama-13B, 1-shot · Translation | accuracy: 80.67 -> 80.13 | No clear effect |
| use task vector instead of ICL on bidirectional bijection (To Upper) | Llama-7B · bijection X<->Y | accuracy: 1.00 -> 0.55 | Hurt |
| use task vector instead of ICL on bidirectional English-French translation | Llama-7B · bijection X<->Y | accuracy: 0.54 -> 0.35 | Hurt |
| use task vector instead of ICL on the copy task | Llama-7B · Copy (x=y) | accuracy: 1.00 -> 0.98 | No clear effect |
| add one task vector to zero-shot prompts | Llama-13B · average ICL accuracy | accuracy: 10.28 -> 72.26 | Improved |
| use triplet-formatted instead of pairwise prompts (ablation) | linear transformer · synthetic linear regression ICL risk | ICL risk | No clear effect |
Closest papers by meaning
- Distributional Alignment as a Criterion for Designing Task Vectors in In-Context Learning (NeurIPS 2026 · similarity 0.92)
- Emergence and Effectiveness of Task Vectors in In-Context Learning: An Encoder Decoder Perspective (ICML 2025 · similarity 0.91)
- Do different prompting methods yield a common task representation in language models? (NeurIPS 2025 · similarity 0.91)
- Capturing In-Context Learning Dynamics with Task Operators (NeurIPS 2026 · similarity 0.90)
- Task Vector Geometry Underlies Dual Modes of Task Inference in Transformers (NeurIPS 2026 · similarity 0.90)
- DETAIL: Task DEmonsTration Attribution for Interpretable In-context Learning (NeurIPS 2024 · similarity 0.90)
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.
