cma-shuffled-label-prompt-isolation

IN premise — summaries/2026/08/24/todd-2023-function-vectors-s0-abstract-chunk-1.md

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

The Causal Indirect Effect metric isolates causal head contribution by replacing head activations with task-conditioned means on shuffled-label prompts, ruling out heads that merely correlate with correct answers.

Summary

This metric checks whether a component truly drives task performance or simply happens to activate whenever the answer is correct, by testing it on deliberately scrambled inputs where that correlation is broken. It matters because without this check, the system could mistake a bystander for a cause and build its understanding on spurious associations rather than real mechanisms.