Applicability of the blind-spot principle to query-key gates

Determine whether the blind-spot principle—that a filter is vulnerable to interference aligned with its own decision variable—also applies to gates operating on the query and key pathways of attention.

Background

The paper establishes that both studied value gates have characteristic blind spots: the norm gate is vulnerable to sufficiently large interference, while the projection gate is vulnerable to interference aligned with its learned projection direction. The authors formulate this as a general blind-spot principle but test it only for value-path gates. Whether the same principle governs gates applied to queries or keys remains unresolved.