Back to archive
#ai#papers#aigen#llm

Interchange ablation

You suspect that a model connection carries information about the book's recipient. Replacing its value with zero may create a situation unlike normal computation. You want to change the information in a more controlled way.

Interchange ablation replaces a selected activation — an intermediate computation's value — with a value from another matched input. Instead of completely removing the signal, we supply one corresponding to another version of the situation.

The model reads “John gave Mary a book”, but one connection receives a value from a sentence where John is the recipient. If the answer changes in line with the substituted information, this supports the hypothesis about the connection's role. This illustrates an intervention, rather than independently proving the full mechanism.

Circuit Condensation, §2, uses this rule for testing circuits. Matching the second input matters: an arbitrary replacement can change many things at once. The result depends on the method — a circuit faithful under these substitutions need not behave identically after zeroing or replacement with a mean.