Persona Dosing: Calibrated Activation Steering for Graded Trait Control
Zehao Jin, Junran Wang, Ruixuan Deng, Jiahao Chen, Jingyuan Zhang, Yuxuan Zhang, Xinjie Shen
Georgia Institute of Technology · Zhejiang University · University of British Columbia
Published September 28, 2026 · on the feed October 2, 2026 · ▲ 51 · arXiv 2609.36388
The 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.
The Brief
2:08A short podcast summary of the paper. Written and voiced by AI.
Transcript
Persona Dosing. Calibrated Activation Steering for Graded Trait Control. By Zehao Jin and six others, from the Georgia Institute of Technology, Zhejiang University, and one other institution. Published on the twenty-eighth of September. No code repository is linked to it.
Activation steering has a knob, the coefficient that sets how hard the intervention pushes. But that knob is not measured in behavior. If you want a model to express a persona to a particular degree, a coefficient does not tell you what degree you will get. Asking for a dose needs a behavioral scale.
Persona dosing is the task the authors define: control a language model with a trait description and a requested mean intensity. Their system, PersonaDose, takes a shared FLAS controller that is conditioned on descriptions, specializes it on persona responses, and then calibrates its flow time against measured trait expression. Calibration is what turns the knob into a scale.
One detail matters for how it is trained. The training responses are never paired with requested target intensities. The mapping from a request to a setting comes from calibration against measurement, not from labelled examples of each level.
The models are Llama-3.1-8B, Qwen3-8B, and Gemma-3-4B. Held at the Persona Vectors coherence floor of seventy-five, PersonaDose raises core-trait expression over contrastive activation addition by thirty-three point two points, eighteen point three points, and seventeen point eight points respectively. On held-out questions, calibration-selected settings keep an expression advantage, although the coherence floor does not hold for every trait there.
Then the targeting itself. Across seven trained traits, calibrated requests miss by a mean of four point seven to six point two points. That accuracy is measured over the targets calibration can actually reach, which is between fourteen and twenty-two of the twenty-eight per model.
Two quantities come apart here. One is how wide a behavioral range the controller has learned. The other is how accurately a request lands inside that range. PersonaDose reports them separately, and the gap between reachable and requested targets is part of the result.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Every quest starts with a wall. Here it is a knob whose numbers mean nothing in behavior.
You study how language models express traits like sycophancy, and you need the same model to show a trait at several measured intensities, while staying readable.
The road ahead
- 1 · The problem ~7 min You can push a model's internal states with a strength knob, but the knob's number does not tell you how strongly a trait will show.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Teach the controller what the trait looks like, then measure how its strength setting maps to trait scores and read that map backwards.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min One shared controller is specialized on persona responses, its flow time is calibrated against judged expression, and requests are translated into flow times.
- 6 · The experiment ~6 min You’ll guess: Held at the same coherence floor, by how many points did PersonaDose raise core-trait expression over contrastive activation addition on Llama-3.1-8B?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.