PersonaDose controls language model traits with calibrated dosing
Persona Dosing: Calibrated Activation Steering for Graded Trait Control
Artificial Intelligence
Summary
Controlling how strongly a language model shows a certain personality trait is tricky because you need a scale to measure it. The authors introduce PersonaDose, a method that teaches a controller to adjust a model’s responses to express traits at different intensities, even when training doesn’t have exact strength labels. They tested this on several models and traits, showing that PersonaDose improves control accuracy and trait expression range compared to previous methods. This lets users choose how much personality a model shows with better precision.
What this means in practice
- •For chatbot developers: Calibrate chatbots to express personality traits more precisely by controlling trait intensity with PersonaDose.
- •For virtual assistant engineers: Adjust virtual assistants’ conversational style on demand by tuning trait levels for natural and consistent persona expression.
Authors
Zehao Jin, Junran Wang, Ruixuan Deng, Jiahao Chen, Jingyuan Zhang, Yuxuan Zhang, Xinjie Shen
Abstract
An activation-steering coefficient sets intervention strength, but requesting a particular degree of persona expression requires a behavioral scale. We study persona dosing: controlling a language model through a trait description and a requested mean intensity. PersonaDose specializes a shared, description-conditioned FLAS controller on persona responses, then calibrates its flow time against measured trait expression. Training responses are not paired with requested target intensities. Across Llama-3.1-8B, Qwen3-8B, and Gemma-3-4B, PersonaDose raises core-trait expression at the Persona Vectors coherence floor of 75 by 33.2, 18.3, and 17.8 points over contrastive activation addition. Calibration-selected settings retain an expression advantage on held-out questions, although the coherence floor does not hold for every trait there. Across seven trained traits, calibrated requests yield mean targeting errors of 4.7-6.2 points over 14-22 calibration-reachable targets out of 28 per model. These results separate the behavioral range learned by a controller from the accuracy of requests within that range.