Skip to content

Criterion E3 — Cross-Task Generalization

Section titled “Criterion E3 — Cross-Task Generalization”
Validity typeExternal
Pass conditionThe mechanism transfers to related tasks, not only the task it was discovered on
Evidence familyBehavioral, Causal
Minimum reportingRelated tasks tested, effect sizes on each, similarity metric between tasks
Common failure modeClaiming generality from a single task

Cross-task generalization asks whether the mechanism operates beyond the specific task used to discover it. An “induction head” found on copying tasks should also operate on other in-context learning tasks if the claim is that it implements general copying.

Satisfied when:

  1. Related tasks are identified and tested. Tasks that share the hypothesized function but differ in surface form.
  2. The mechanism operates on those tasks. Ablation or patching shows the mechanism is causally involved in the related tasks.
  3. The scope is bounded. The claim states which tasks the mechanism covers and which it does not.

The greater-than circuit was discovered on “The war lasted from 1723 to 17__” prompts. Cross-task generalization asks whether the same circuit handles other ordinal comparisons — months, alphabet position, numbered lists. If it only works on year comparisons, the mechanism is narrower than “greater-than” implies.

Required for Validated tier. Without cross-task generalization, a mechanism labeled with a general function word (“greater-than,” “copying,” “refusal”) may be specific to one task format. The label imports scope the evidence has not established (see V4, unlicensed labeling).