Linear Task Vector method matches in-context learning at lower cost, paper reports
A new arXiv paper proposes a metric for judging task vectors and a closed-form method that reportedly beats existing baselines on accuracy and speed.
In plain words
Large language models can adapt to new tasks by reading examples in their prompt, but long prompts make inference slow and costly. The paper's authors compress those examples into a compact hidden-state representation called a task vector. They propose a way to measure how closely such a vector reproduces the model's next-token predictions, then use it to build a method they say is more accurate and faster.
What was announced
The authors argue that task vectors, which compress demonstrations into hidden states, have been judged only by downstream accuracy, giving little guidance for designing better extraction methods. They introduce d_NTP, a metric measuring the discrepancy in next-token probabilities between task-vector-based and in-context inference, and report it acts as a performance proxy with a strong negative correlation with accuracy. Their Linear Task Vector method minimises this metric using a closed-form linear mapping that estimates demonstration effects through regression.
Across eight classification benchmarks and five LLMs, the authors report that LTV consistently outperformed existing task vector baselines, improving average accuracy by 9.2% while reducing inference latency, and also beat baselines on regression tasks. They further claim that task vectors extracted from a larger model improved a smaller model's performance by 6.4%. These figures come from the paper's own experiments and have not been independently verified.
What it does not show
- The source is the paper's abstract only; full experimental details, model identities and benchmark names are not available in the supplied excerpt.
- All accuracy, latency and transferability figures are the authors' own claims from the abstract and have not been independently confirmed.
For security and IT teams
- Status
- The work is a v2 arXiv preprint with an 'replace' announce type, so the paper may be revised and is not a peer-reviewed publication.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.