EvoLib: Turning experience into evolving knowledge

Microsoft · 2 months ago · 21

Microsoft Research introduced EvoLib, a framework that enables large language models to extract reusable skills and insights from their own experiences during inference without model updates or external labels. The system consolidates and reweights knowledge over time, allowing it to convert test-time compute into performance improvements more effectively than existing memory-based approaches across mathematical reasoning, coding, and decision-making tasks. This enables AI systems to continually learn and improve after deployment by building evolving libraries of transferable knowledge rather than accumulating static memories.