跳到正文
arXiv cs.AI· Jihoon Kwon, Jiwon Choi, Jy-yong Sohn·· 4 小时前AI 评分31

分布对齐作为上下文学习任务向量设计准则:LTV 方法提升 9.2% 准确率

Distributional Alignment as a Criterion for Designing Task Vectors in In-Context Learning

AI 导读

研究者提出以分布对齐为准则设计上下文学习(ICL)的任务向量,并引入 $d_{\text{NTP}}$ 指标衡量任务向量推理与 ICL 推理在下一 token 概率上的差异,发现该指标与下游准确率呈强负相关。

正文 · 原文

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. While task vectors offer a promising alternative by compressing demonstrations into compact hidden-state representations, their quality has been evaluated only through downstream task accuracy. This indirect criterion provides limited insight into how to design more effective task vector extraction methods. In this paper, we posit that inference using task vectors should align their predictive distribution with that of ICL. To quantify this, we introduce $d_{\text{NTP}}$, a metric that measures the discrepancy in next-token probabilities between task vector-based and ICL-based inference. Our empirical analysis reveals that $d_{\text{NTP}}$ serves as a performance proxy, exhibiting a strong negative correlation with downstream accuracy. Motivated by this, we develop Linear Task Vector (LTV), a method designed to minimize $d_{\text{NTP}}$ via a closed-form linear mapping that estimates demonstration effects through regression. Across eight classification benchmarks and five LLMs, LTV consistently outperforms existing task vector baselines, improving average accuracy by 9.2\% while reducing inference latency. We further show that LTV outperforms the baselines on regression tasks. Moreover, we investigate the transferability of LTV across different model scales; an aspect that has remained nascent in task vector research. Specifically, we empirically show that task vectors from a larger model can enhance a smaller model's performance by 6.4\%, suggesting a new utility for extracted task representations.
Comments: Accepted at NeurIPS 2026. 29 pages
Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
Cite as: arXiv:2605.20730 [cs.CL]
  (or arXiv:2605.20730v2 [cs.CL] for this version)
  https://doi.org/10.48550/arXiv.2605.20730

arXiv-issued DOI via DataCite

Submission history

From: Jihoon Kwon [view email]
[v1] Wed, 20 May 2026 05:26:38 UTC (2,391 KB)
[v2] Tue, 17 Nov 2026 05:53:04 UTC (2,385 KB)

来源:arXiv cs.AI · arxiv.org