AdaptFuse: Training-Free Sequential Preference Inference via Externalized Bayesian Computation
AdaptFuse is a training-free method that allows Large Language Models to learn user preferences sequentially without fine-tuning or storing sensitive data. It separates probabilistic computation into a symbolic module tracking Bayesian beliefs and a frozen LLM providing semantic reasoning. This combination, using entropy-adaptive fusion, consistently outperforms other methods by letting the symbolic model take over as interaction evidence accumulates.