AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Task Vectors, Learned Not Extracted: Performance Gains and Mechanistic Insights

Haolin Yang, Hakaze Cho, Kaize Ding, Naoya Inoue

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Task vectors and steering

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 changedWhereMeasuredEffect
train the task vector by gradient descent instead of extracting ittask vector injection · 7-dataset avg, layer 16 last tokenaccuracy (%): 37.80 (Vanilla TV), 37.30 (FV) -> 83.49Improved
inject LTV at a different token positionLTV · 7-dataset avgaccuracy (%): 83.49 -> 78.39No clear effect
inject at multiple positions (ablation)LTV · 7-dataset avgaccuracy (%): 83.49 -> 86.43Improved
inject at multiple positions (ablation)Function Vector · 7-dataset avgaccuracy (%): 37.30 -> 31.88Hurt
inject at many layers and positions simultaneously (ablation)LTV · 7-dataset avgaccuracy (%): 83.49 -> 51.39Hurt
inject LTV into 8-shot ICL promptsLTV · 7-dataset avgaccuracy (%): ICL slightly below 80 -> 84.61Improved
use LTV instead of Prefix Tuning or LoRA as PEFTLlama3.1-8B · SST-2accuracy (%): 85.67 (Prefix), 91.63 (LoRA) -> 92.89Improved
reconstruct TV effect via attention-head OV circuits (ablation)LTV · 7-dataset avgaccuracyImproved
ablate top-10% key heads vs random 10% heads (ablation)LTV · 7-dataset avgaccuracyHurt
apply SNLI-trained LTV to other tasksLTV · cross-taskaccuracy (%): 46.21 on RTE, 0.00 on SST-2/TRECMixed

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy