fv-composition-outperforms-10shot-icl

IN premise — summaries/2026/08/24/todd-2023-function-vectors-s0-abstract-chunk-2.md

Created 2026-08-25T02:58:40+00:00

Composed FVs (v*_BD = v_B + v_D) outperform 10-shot ICL on some tasks: Last-Country-Capital achieves 0.60 ± 0.02 accuracy with composed FV vs 0.32 ± 0.02 with ICL, and Last-Capitalize-First-Letter achieves 0.95 ± 0.00 vs 0.75 ± 0.01.

Summary

Combining two separate task vectors into a single one can beat the standard approach of showing a model ten worked examples, sometimes by a wide margin (nearly double the accuracy on one task). This suggests that vector composition is a more efficient and reliable way to chain or merge capabilities than relying on in-context learning, which has practical implications for how the system assembles multi-step tasks without burning prompts or degrading performance.