Letting Tutor Personas Speak Up for LLMs: Learning Steering Vectors from Dialogue via Preference Optimization
Researchers have developed a method to control the behavior of large language models (LLMs) by learning "steering vectors" from human tutor-student dialogues. This approach allows LLMs to adopt different tutoring personas without explicit prompting, capturing variations in instructional strategies and affective support. The steering vectors improve semantic alignment with desired tutor responses and are evaluated favorably, demonstrating an interpretable way to guide LLM behavior using real-world dialogue data. AI
IMPACT Enables more nuanced and adaptable LLM-driven educational tools by allowing persona customization.