Cross-Model Persona Vector Transfer: Enabling Different Architecture Large Language Models to Share 'Personality'

Core breakthrough: Achieves cross-architecture persona vector transfer, extracting persona feature vectors from Qwen2.5-7B and successfully applying them to GPT-OSS 20B, opening a new path for behavioral control of large language models (LLMs). This technology breaks the representation barriers between models of different architectures, turning personality into an extractable and transferable modular component, which has significant technical and application value.

With the widespread application of LLMs, there is an urgent need to precisely control the model's 'personality/behavioral style' (e.g., friendly customer service, rigorous education). Traditional methods (limited effectiveness of prompt engineering, high fine-tuning costs) have shortcomings. The concept of 'persona vectors' has emerged, but the feasibility of cross-architecture transfer is a key unsolved problem.

Core breakthrough of the cross-model-persona-steering project: Extracting persona vectors from Alibaba's Qwen2.5-7B and successfully applying them to OpenAI's GPT-OSS 20B. Although the two have different architectures, training data, and alignment strategies, the study shows that there exists a high-level semantic space of 'universal personality language'.

The core process consists of three steps: 1. Extraction: Analyze the personality-related activation patterns of the source model (Qwen2.5-7B) and separate the vector direction representing a specific personality; 2. Alignment: Resolve cross-architecture spatial differences through linear transformation, contrastive learning, and intermediate layer selection; 3. Injection: Inject the aligned vector into the forward propagation of the target model (GPT-OSS 20B) and adjust activation values to achieve personality transfer.

Experiments show that the injected GPT-OSS 20B has high personality consistency while retaining core capabilities/knowledge. Evaluation metrics include: personality consistency (matching degree with target description), capability retention (standard benchmark performance), stability (effect in different scenarios), and controllability (adjusting personality intensity via vector strength).

1. Efficient customization: Iterate personality design on small models and transfer to large models, reducing costs; 2. Standardized personality library: Similar to a font library, users can choose to apply it to models that support transfer; 3. Multi-model consistency: Ensure unified personality across all models in complex systems; 4. Safety alignment: Provide tools for AI safety research to explore the relationship between personality and behavior.

Limitations: Transfer accuracy is not 100%, extreme architectures are difficult to align, and complex personalities are hard to represent with a single vector. Future directions: Explore nonlinear alignment methods, fine-grained personality decomposition techniques, and establish a theoretical foundation for transfer.

This project marks an important progress in LLM behavioral control, turning personality from an inherent attribute into a pluggable module. This paradigm shift not only has technical significance but also triggers thinking about the essence of AI: When personality can be transferred, how to define the relationship with AI?