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

Name: How Alignment Routes: Localizing, Scaling, and Controlling Policy Circuits in Language Models
Uploaded: 2026-10-22T12:48:07.429978+00:00
Duration: 452 s
Description: Researchers found that policy routing in language models depends on a specific 'gate-amplifier mechanism' localized at layers 15–16. They discovered a gate head (L17.H17) that acts as a sparse attention entry point, reading detection signals and routing them toward refusal. This circuit can be controlled by modulating the detection layer, suggesting precise policy steering is possible.

Researchers found that policy routing in language models depends on a specific 'gate-amplifier mechanism' localized at layers 15–16. They discovered a gate head (L17.H17) that acts as a sparse attention entry point, reading detection signals and routing them toward refusal. This circuit can be controlled by modulating the detection layer, suggesting precise policy steering is possible.

Open the full episode