Online Experiential Learning for Language Models: Self-Improvement from Real Interactions Without Offline Retraining
arXiv 2603.16856·medium signal
This paper challenges the dominant offline training paradigm by proposing online experiential learning — LLMs continuously improve from real-time interaction signals without requiring offline annotation or retraining cycles. The approach leverages the rich experiential feedback available during deployment that current methods leave underutilized. Directly relevant to production agent deployments where models need to adapt to specific environments and user patterns over time.