Distributional Alignment as a Criterion for Designing Task Vectors in In-Context Learning
The paper discusses a new method for designing task vectors in in-context learning (ICL) that aims to improve efficiency and accuracy. It introduces a metric called $d_{\text{NTP}}$ to measure the alignment of predictive distributions between task vector-based and ICL-based inference. The proposed Linear Task Vector (LTV) method shows significant improvements in accuracy and reduced inference latency across various benchmarks.
- ▪In-context learning allows large language models to adapt to new tasks but faces challenges with increasing inference costs.
- ▪The authors propose a new metric, $d_{\text{NTP}}$, to evaluate the effectiveness of task vector extraction methods.
- ▪The Linear Task Vector (LTV) method developed in the study improves average accuracy by 9.2% while reducing inference latency.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Computation and Language arXiv:2605.20730 (cs) [Submitted on 20 May 2026] Title:Distributional Alignment as a Criterion for Designing Task Vectors in In-Context Learning Authors:Jihoon Kwon, Jiwon Choi, Jy-yong Sohn View a PDF of the paper titled Distributional Alignment as a Criterion for Designing Task Vectors in In-Context Learning, by Jihoon Kwon and 2 other authors View PDF HTML (experimental) Abstract:In-context learning (ICL) allows large language models (LLMs) to adapt to new tasks through demonstrations, yet it suffers from escalating inference costs as context length increases.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
