fv-extraction-uses-100-clean-prompts

IN premise — summaries/2026/08/24/todd-2023-function-vectors-s2-function-vector-causal-effects-cannot-be-recovered-from-the--chunk-1.md

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

FV extraction (Eq. 5) computes v_t as the sum of mean task-conditioned head activations over |P_t| = 100 clean 10-shot prompts, with heads selected by ranking average indirect effect (AIE) over 25 corrupted 10-shot prompts per task.

Summary

The system identifies which internal attention heads are task-relevant by stress-testing them against 25 corrupted examples, then reads out the average behavior of only those selected heads across 100 clean examples to produce a single numerical signal. This fixed recipe is the measurement protocol that every downstream step depends on, so any change to those prompt counts or the selection method would alter every value the system reports.