AI
Persona Dosing: Calibrated Activation Steering for Graded Trait Control
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 27/09/2026, 20:00:00 UTC
Briefing created: 03/10/2026, 08:20:36 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 02/10/2026, 19:18:24 UTC.
Briefings
PersonaDose is a system that calibrates language model activation steering to control trait expression via behavioral intensity. The study tested this method across Llama-3.1-8B, Qwen3-8B, and Gemma-3-4B models. Results show that while calibrated requests achieve mean targeting errors of 4.7-6.2 points, the system successfully raises core-trait expression at the Persona Vectors coherence floor by 17.8 to 33.2 points over contrastive activation addition. The method separates learned behavioral ranges from request accuracy, though calibration settings do not consistently maintain expression advantages across all traits.
Key points
- PersonaDose calibrates language model activation steering to control trait expression through behavioral intensity.
- Across Llama-3.1-8B, Qwen3-8B, and Gemma-3-4B, the system raises core-trait expression by 17.8 to 33.2 points over contrastive activation addition.
Context and limitations
Training responses were not paired with requested target intensities, and calibration-selected settings do not consistently maintain expression advantages across all trained traits.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 03/10/2026, 08:20:36 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeibrodmq2uiaqwz64gpuyzxjrdhhodo4tajysjnay33hrjc5yp5zfe