Interchange Interventions & DAS

Activation patching copies a vector at a chosen site. Interchange interventions make a more targeted claim: there is a subspace whose values act like a specific causal variable. Distributed alignment search (DAS) trains a rotation, chooses a low-rank slice of the rotated activation, and swaps only that slice between two prompts while the model itself stays frozen.

The figures on this page are schematic: the percentages illustrate how the method behaves, not measured results from a specific model.

1. What gets trained

DAS does not train the language model. It trains an alignment around one layer: rotate the activation, expose a small number of rotated coordinates, and optimize that slice so swapping it transfers the target variable. The downstream model is then run normally. If the output flips under the learned slice but not under a random slice of the same rank, the evidence is stronger than plain decodability.

2. Floor, ceiling, and transfer index

The raw flip rate is hard to read by itself. A corrupted prompt may already flip some of the time, and even a perfect intervention may not reach 100 percent. The useful number is normalized between a floor and a ceiling: how much of the available behavioral transfer did the learned interchange recover?

3. Expressive fit is not yet mechanism

A trained alignment can be too expressive. If the aligner finds a subspace that predictably moves the output but is off the path the model actually uses, it has shown a useful control surface rather than the internal mechanism. The critique is not that DAS is invalid; it is that an efficacy number needs supporting evidence about location, controls, and path specificity.

The practical reading is conservative. DAS can turn a probe-like readout into a causal interchange test. It does not, by itself, prove that the model stores the variable in exactly the learned coordinates or that this is the only path used for the behavior.