Metraipolis › archive › dc_1e896ec32273

Persona Dosing: Calibrated Activation Steering for Graded Trait Control — arXiv:2609.36388

Source
arXiv (en)
Original
https://arxiv.org/abs/2609.36388
License (as archived)
CC0-1.0
Archived
2026-10-07T16:22:33.962Z
SHA-256 of the archived text
906300a861988ec1d10e27f66a426298e6ef565ea6224fc2818a46baedb10043
Stable URL of this page
https://metraipolis.com/documento/dc_1e896ec32273

How to cite

Persona Dosing: Calibrated Activation Steering for Graded Trait Control — arXiv:2609.36388 — arXiv (en), https://arxiv.org/abs/2609.36388 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_1e896ec32273 (Metraipolis archive dc_1e896ec32273, sha256 906300a861988ec1d10e27f66a426298e6ef565ea6224fc2818a46baedb10043)
BibTeX
@misc{dc1e896ec32273,
  title = {Persona Dosing: Calibrated Activation Steering for Graded Trait Control — arXiv:2609.36388},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2609.36388},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_1e896ec32273; sha256 906300a861988ec1d10e27f66a426298e6ef565ea6224fc2818a46baedb10043},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.36388. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

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.

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-30 in cs.AI).
Categorie arXiv: Artificial Intelligence (cs.AI).
Pagina del lavoro: https://arxiv.org/abs/2609.36388 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito).
Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina.
Letto il 2026-10-07T15:54:53.495Z dalla pagina https://arxiv.org/catchup/cs.AI/2026-09-30?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).