Noesion Radarnoesion.ai
Research2026-08-03

Microsoft Research released EvoLib, the official code for the paper "Test-Time Learning with an Evolving Library" (arXiv:2605.14477) by Weijia Xu, Alessandro Sordoni, Chandan Singh, Zelalem Gero, Michel Galley, Xingdi Yuan and Jianfeng Gao. EvoLib lets black-box LLMs accumulate and reuse knowledge across problem instances without parameter updates or ground-truth supervision, maintaining a library of modular skills and reflective insights extracted from the model's own inference trajectories, weighted by Information Gain and Future IG and periodically consolidated. Reported cost-performance curves show higher accuracy than Best-of-N, RSA and Dynamic Cheatsheet on BigCodeBench Hard (GPT-4o), LiveCodeBench v6 Hard and HMMT 2025-2026 (o4-mini), plus AgentBoard ScienceWorld/PDDL agentic tasks. The MIT-licensed release uses Azure OpenAI endpoints and is stated to be for research purposes only, not recommended for commercial or high-risk use.

Send this to someone who needs it

Shares the story and its sources. Nothing about you.

What does this mean for your job?

This is the story as everyone gets it. Once a week we send you the version written for your role — what changed, why it matters for the work you actually do, and one thing to try. Free while we tune it.