Improve Large Language Model Systems with User Logs

The article introduces UNO (User log-driveN Optimization), a framework designed to enhance large language model systems by leveraging user interaction logs for continual learning. UNO processes these logs into semi-structured rules and preference pairs, utilizes query-and-feedback-driven clustering to address data heterogeneity, and assesses the cognitive gap between the model's prior knowledge and user feedback to filter out noise. Experimental results demonstrate that UNO significantly outperforms existing methods like Retrieval Augmented Generation (RAG) and memory-based approaches, highlighting its potential for improving LLM performance in real-world applications.