New 'PsychAdapter' Tailors LLMs to User Personality
A new technique called "PsychAdapter" enables large language models to adapt their dialogue and recommendations to reflect a user's personality, traits, and even mental state. For AI tutors, this could mean dynamically adjusting tone and encouragement for an anxious or struggling young reader, creating a more supportive learning experience.
PsychAdapter modifies the core transformer architecture, allowing psychological traits to influence every layer of the model. Researchers have successfully applied this lightweight adaptation to models including Google's Gemma, Meta's Llama 3, and OpenAI's GPT-2. The technique moves beyond simple prompting by using continuous scores for traits like the Big Five personality dimensions (openness, extraversion, neuroticism) and mental health variables like depression. This allows for more nuanced control compared to conditioning models on discrete attributes like age or gender. In evaluations by expert human raters, the generated text matched the intended Big Five personality traits with an average accuracy of 87.3%. For mental health states like depression and life satisfaction, the accuracy reached 96.7%. This research was a collaboration involving engineers and psychologists from Stanford, Stony Brook University, and the University of Pennsylvania, among others. For an adaptive learning system, this method could directly inform the reward function in a reinforcement learning loop. An AI tutor could adjust its level of extraversion or agreeableness in real-time based on tracking a student's frustration or disengagement, creating a more motivating interaction model. The model was trained on a large corpus of social media posts and blogs, with psychological scores estimated from the language used in those texts. This approach suggests a pathway for edtech platforms to develop their own datasets to fine-tune models for specific student populations and learning contexts. Unlike traditional adaptive systems that might use a multi-armed bandit approach to select the next piece of content, PsychAdapter can modify the delivery of that content. For a K-3 reader, this means the AI could generate praise that is not just algorithmically triggered but also tonally matched to the child's inferred emotional state.
