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.
More from coding & agent
- Four Gemini 4 Argon agents beat Minecraft's Ender Dragon in 27 hours with 2,783 self-written programs — robertarail · 2026-10-16
- Dev builds oddly satisfying self-fixing sand component that repairs itself — prasenx · 2026-10-16
- When Does an Agent Actually Need an API Gateway? Reddit Weighs Direct Calls vs. Unified Endpoints — TurbulentStiffness · 2026-10-16
- optimAIzr: Open-Source Local CLI Tracks Where Your Claude Code and Codex Tokens Go — stichstichstich · 2026-10-16
- Local tool calling compared: how Gemma 4, Llama 3 and Mistral differ — leslysandra · 2026-10-16
- Harness, a command center for coding agents, hits #2 on Product Hunt — dee_hw · 2026-10-16