sae-attribution-linear-formula

Summary

This describes the exact recipe for measuring how much each SAE feature pushes the model's output in a given direction: you multiply the feature's activation strength by its direction vector, then take the dot product with how sensitive the output is to changes at that point. The choice of zero as the baseline (rather than comparing against a second example) means attribution scores here are absolute contributions from each feature, not differences between two situations, which affects how you interpret and compare them across prompts.