Skip to content
Validity typeExternal
Pass conditionThe mechanism operates on diverse prompts, not only the specific examples used for discovery
Evidence familyBehavioral
Minimum reportingNumber and diversity of test prompts, effect size distribution across prompts
Common failure modeDiscovering and testing on the same small prompt set

Prompt generalization asks whether the mechanism works beyond the specific inputs that were used to find it. A circuit discovered on 20 IOI sentences and tested on the same 20 sentences has not demonstrated generalization.

Satisfied when:

  1. Test prompts differ from discovery prompts. The evaluation set is distinct from the set used to identify the mechanism.
  2. Prompts are diverse. They vary in length, structure, and lexical content, not just token substitution.
  3. The effect holds across the distribution. The mechanism’s effect size is consistent, not driven by a few outlier prompts.

IOI circuit studies typically use template-generated sentences (“When Mary and John went to the store, John gave a drink to ___”). The circuit is discovered and tested on variants of this template. Prompt generalization requires testing on naturally occurring indirect-object constructions with different syntactic structures — relative clauses, passives, embedded contexts — not just slot-filled versions of one template.

Required for Triangulated where the claim does not assert reach beyond the systems tested; cross-model generalization (E4) carries the same gate where it does. Without prompt generalization, the mechanism may be an artifact of the specific input distribution used for discovery, not a general property of the model.