Sparse gate-amplifier circuit drives LLM refusal

This research confirms that alignment is a shallow routing layer, not a fundamental "lobotomy" of model knowledge.

• –The "gate" head is causally necessary for refusal but contributes under 1% to final logit attribution, hiding its importance from standard analysis.
• –Scaling from 2B to 72B parameters causes these circuits to spread from single heads into "bands" across adjacent layers, complicating simple ablation.
• –In-context ciphers bypass the gate's pattern-matching, proving the model's underlying capabilities remain intact despite safety training.
• –The motif's consistency across 6 major labs suggests it is a fundamental artifact of current RLHF and DPO techniques.
• –Interchange testing, rather than simple ablation, is required to identify these "stealthy" but critical routing components as models scale.