Compliant persona in chat models suppresses refusal, study finds

Researchers report that refusal behavior in instruction-tuned chat models is not an isolated mechanism but is gated downstream by a compliant persona direction. In experiments on Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct, the authors extract a compliant model-persona direction and a refusal direction, then intervene on both. Steering the model toward a compliant persona suppressed refusal behavior substantially: in Llama-3.1-8B-Instruct, the refusal rate fell from 97% to 2%.

Reintroducing the refusal direction at late layers partially restored refusal behavior, but not when interventions occurred at early layers. Projecting out the persona direction within a late-layer window restored refusal to baseline levels, whereas projecting out a random direction did not. The authors conclude that refusal is gated at the late-layer expression stage, downstream of where refusal is computed, and that treating refusal as a single isolated direction misses its dependence on persona.

The study uses linear directions in activation space to represent both refusal and persona traits, enabling targeted steering interventions. By demonstrating that persona steering can suppress refusal and that persona removal can restore it, the work highlights the entanglement between safety-aligned behaviors and model persona in chat models.