Persona Dosing: Calibrated Activation Steering for Graded Trait Control
The paper tests whether persona traits can be dialed to requested intensity levels rather than just switched on.
PersonaDose calibrates an activation-steering controller against measured trait expression, without training on responses paired to target intensities. On Llama-3.1-8B, Qwen3-8B, and Gemma-3-4B, it improves core-trait expression over contrastive activation addition at the stated coherence floor. The authors report calibration still helps on held-out questions, though the coherence floor does not hold for every trait. Across seven trained traits, calibrated requests land within 4.7–6.2 points on reachable targets. HF Daily Papers' note
PersonaDose calibrates an activation-steering controller against measured trait expression, without training on responses paired to target intensities. On Llama-3.1-8B, Qwen3-8B, and Gemma-3-4B, it improves core-trait expression over contrastive activation addition at the stated coherence floor. The authors report calibration still helps on held-out questions, though the coherence floor does not hold for every trait. Across seven trained traits, calibrated requests land within 4.7–6.2 points on reachable targets. HF Daily Papers' note
score 4