Psy2Vec: Zero Fine-tuning Persona Vector Learning with Differentiable Soft-Label Gradient Descent and User-Defined Psychometric Scales

Role Persona Injection is critical for improving role consistency in large language models. However, existing vector-based steering methods—such as Persona Vectors and Linear Personality—rely on statistical mean differences or linear regression, which operate at a coarse semantic granularity and struggle to precisely calibrate "multiple-choice distributional responses"; moreover, the optimal steering layer is often selected heuristically. To address these issues, we propose Psy2Vec, Zero Fine-tuning Persona Vector Learning with Differentiable Soft-Label Gradient Descent and User-Defined Psychometric Scales. First, we adopt a configurable "category–dimension–item" structured scale, which allows dimensions and items to be freely added or removed. Then, we employ a Tutor Model to output option percentages, forming soft-labels carried by the character's option probability distribution. Next, through Computed Optimal Steering Layer Selection, we compare the activation gaps between high- and low-persona conditions layer by layer to derive the optimal steering layer, thereby replacing heuristic selection. Finally, we optimize the persona vector via differentiable RMSE gradient descent, achieving zero fine-tuning throughout without modifying any model weights. Preliminary experiments comparing both approaches on the same questionnaire demonstrate that Psy2Vec achieves lower persona-consistency error.