Microsoft's EvoLib Enables LLMs to Evolve Knowledge During Inference

Microsoft Research · rss · 2026-07-31

Microsoft Research introduced EvoLib, a framework enabling LLMs to learn continuously from their own experience during inference without updating model parameters or requiring ground-truth labels.

Core Mechanisms

Knowledge Transformation: Distills past attempts into reusable skills and reflective insights rather than merely archiving raw memories.
Dynamic Evolution: Consolidates new and old knowledge into generalized rules, and dynamically reweights knowledge based on long-term utility to prune ineffective memories.

Key Results

Across math reasoning, constrained coding, and long-horizon decision-making tasks, EvoLib consistently outperforms traditional retrieval-based memory methods while using compute more efficiently. It also demonstrates strong robustness to random task ordering, making it highly practical for real-world scenarios with mixed user requests.

Original post →

Microsoft's EvoLib Enables LLMs to Evolve Knowledge During Inference

More from coding & agent