EvoLib: Turning Experience into Evolving Knowledge

Microsoft Research introduces EvoLib, a self-supervised approach that lets large language models learn from their own experience during inference. It does not require ground-truth labels, external feedback, or model updates. Instead, EvoLib converts prior attempts into reusable skills and reflective insights, then continually refines, consolidates, and reweights that knowledge so it can transfer across tasks. The approach is designed to work with black-box language models and API-deployed AI systems, extending memory from simple storage and retrieval toward an evolving learning mechanism.