Why it is a question
THEORY-282 holds the construction. A patch along a direction that mixes a component the next layer cannot read with one that is constant on the data moves the output as changing the variable would. Neither part alone has any effect. Makelov, Lange and Nanda then measure the same signature in GPT-2 Small’s layer-8 MLP on indirect-object identification: 46.7% FLDD, falling to 13.5% without the null-space part. They find it again in GPT2-XL’s MLPs on factual recall. They leave open a rival reading: two subspaces in the same component that both mediate the feature and cancel on the data. The reply by Wu et al. contests the label, not the measurements. Until 2026-10-20 this was THEORY-282’s promote_when.
What would answer it
- A replication, in a second model family, of the dependence on the disconnected component, for a feature whose mediating location is known independently.
- A test that moves the two candidate subspaces while holding the component’s total output fixed on the data, which separates the dormant-direction reading from cancellation.
Against it: interventions that stay on the data distribution and show that the component the illusory direction lives in does take part in the behaviour.
What would not answer it: more successful patches, or more directions that fail the rowspace test, since the construction predicts both.