UNO Framework Improves LLMs Using User Logs

A new framework called UNO (User log-driveN Optimization) aims to improve large language model (LLM) systems by leveraging user interaction logs from real-world deployment. As scaling training data and model parameters faces diminishing returns and high-quality data scarcity, researchers are turning to continual learning from user logs, which offer authentic human feedback and procedural knowledge. However, these logs are unstructured and noisy, making it difficult for vanilla LLM systems to extract useful signals. UNO addresses this by first distilling logs into a structured format, enabling effective optimization despite challenges like off-policy problems. The framework is detailed in a paper on arXiv (2602.06470).