How Few-Shot Examples Add Up: A Causal Decomposition of Function Vectors in In-Context Learning

In three linesMechanistic study of in-context learning (ICL): n-shot function vectors decompose linearly into individual example contributions. Models adaptively reweight demonstrations via attention, favoring informative and unambiguous examples. Query-Key alignment dominates function vector quality.

Summary generated by Claude — human-verified