Closure-Validated Circuit Discovery in Consideration Heads: Co-activation Proposes, Ablation Disposes

arXiv:2606.09607v1 Announce Kind: cross
Summary: Interpretability more and more treats teams of parts, not particular person items, as the essential object, and proposes to search out them by clustering co-activation statistics. We ask whether or not such an affordable sign really identifies an attention-head circuit. Adapting a sparse-autoencoder clustering recipe to consideration heads — however validating by causal ablation relatively than reconstruction — we cluster heads after which run a closure take a look at: ablate the found group and examine per-example injury to matched-random controls. Throughout two dense 1B-scale fashions (Pythia 1B, OLMo 1B) and two enter distributions, the communities move closure. In a Combination-of-Specialists mannequin (OLMoE-1B-7B), route-conditional clustering recovers a statistically actual sign that nonetheless doesn’t survive closure — ablation improves loss, the mistaken route. Extending closure throughout coaching, attention-target selectivity and participation ratio decouple from perform in each instructions. We conclude that an affordable sign is a circuit proposal, not a confirmed circuit; closure is what separates them.



Source link