hendel-hypothesis-procedure-accuracy-retention
IN premise — summaries/2026/08/24/hendel-2023-icl-task-vectors-s4-robustness-of-task-vectors.md
Created 2026-08-25T02:58:04+00:00
The Hypothesis procedure (task-vector patching) achieves approximately 80–90% of Regular ICL accuracy, while the Baseline (no demonstrations, no θ) reaches only 10–20%.
Summary
You don't need to show a model full examples to make it perform well on a task. A single compact vector that encodes the task can recover most of the accuracy you'd get from demonstrations, while doing nothing at all leaves the model nearly blind, confirming that the vector itself is carrying the meaningful signal.