EvoLib: Turning experience into evolving knowledge
Microsoft says EvoLib lets language models turn inference-time attempts into reusable skills without retraining the model.
The framework stores distilled skills and reflective lessons rather than raw memory logs.
It consolidates similar lessons and reweights them by usefulness, including their value for future tasks.
Microsoft says EvoLib beat retrieval-based and abstract memory baselines on math, coding, and long-horizon environment tasks while using tokens more efficiently.
The post says code and experiment results are available for further research.
Microsoft Research AI's note
The framework stores distilled skills and reflective lessons rather than raw memory logs.
It consolidates similar lessons and reweights them by usefulness, including their value for future tasks.
Microsoft says EvoLib beat retrieval-based and abstract memory baselines on math, coding, and long-horizon environment tasks while using tokens more efficiently.
The post says code and experiment results are available for further research.
Microsoft Research AI's note
score 5