From the source
Microsoft Research introduces EvoLib, a framework that enables large language models to learn from their own experience during inference by transforming raw experiences into an evolving library of reusable skills and reflective insights, which are continually refined, consolidated, and reweighted over time.
The framework does not require model updates and can be applied to any black-box language model.
Evaluated on mathematical reasoning, code generation, and decision-making tasks, EvoLib outperforms retrieval-based memory approaches and other abstract memory mechanisms with more efficient token usage, and shows robustness to random task order.
