How Alignment Routes: Localizing, Scaling, and Controlling Policy Circuits in Language Models

Researchers identified a sparse routing mechanism in alignment-trained language models where gate attention heads detect content and trigger amplifier heads that boost refusal signals. The study analyzed 9 models from 6 labs and found this routing mechanism distributes at scale while remaining controllable through signal modulation.

• →A consistent sparse routing mechanism was found across 9 language models from 6 different labs for handling policy-violating content.
• →Gate attention heads detect problematic content and trigger downstream amplifier heads that boost refusal signals.
• →The routing mechanism becomes more distributed but weaker as models scale up, with ablation effects up to 17x weaker.
• →Signal modulation allows continuous control of policy strength from hard refusal to factual compliance.
• →Cipher encoding reveals structural separation between intent recognition and policy routing, with the gate head's contribution collapsing by 78% in Phi-4.