Researchers have introduced PersonaDose, a method for controlling a language model's persona through a trait description and a requested mean intensity, instead of manually tuning a raw activation-steering coefficient. The work is described in a paper posted to Hugging Face.
The core idea is that a steering coefficient sets intervention strength, but it does not directly indicate how much of a trait will actually be expressed. PersonaDose addresses this by specializing a shared, description-conditioned FLAS controller on persona responses, then calibrating its flow time against measured trait expression. Notably, training responses are not paired with requested target intensities.
Across three models — Llama-3.1-8B, Qwen3-8B, and Gemma-3-4B — PersonaDose raised core-trait expression at the Persona Vectors coherence floor of 75 by 33.2, 18.3, and 17.8 points respectively over contrastive activation addition.
Calibration-selected settings retained an expression advantage on held-out questions, though the coherence floor did not hold for every trait there. Across seven trained traits, calibrated requests yielded mean targeting errors of 4.7–6.2 points over 14–22 calibration-reachable targets out of 28 per model.
The authors say these results separate the behavioral range learned by a controller from the accuracy of requests within that range. The method builds on FLAS, presented at NeurIPS 2026.