A-EVO-Lab/a-evolve
The official repository of "Position: Agentic Evolution is the Path to Evolving LLMs".
What it solves
A-Evolve 为 AI Agent 的自主自我改进提供了一套通用基础设施。它消除了对手动搭建测试框架(harness engineering)的需求,允许任何 Agent 通过各种进化算法,在不同领域中通过突变其自身的提示词、技能和记忆来实现自我改进。
How it works
该系统将 Agent 可进化的状态——例如系统提示词、技能库和情节记忆(episodic memory)——视为文件系统上的标准目录结构(即 "Agent Workspace")。演化引擎随后运行一个持续循环:它让 Agent 解决任务,观察结果,分析性能日志以突变工作空间文件,并这些突变在通过 git-tagging 提交之前,在保留任务(holdout tasks)上进行验证。
Who it’s for
它专为从事 LLM 自我优化、强化学习和 Agent 架构研究的 AI 研究人员和开发者设计,旨在让他们在无需手动调优的情况下,测试和进化 Agent。
Highlights
- Universal Framework: Supports "Bring Your Own Agent" (BYOA), "Bring Your Own Benchmark" (BYOE), and "Bring Your Own Algorithm" (BYO-Algo).
- Autonomous Post-Training: Capable of post-training large models (up to 550B parameters) with no human in thet loop.
- Built-in Algorithms: Includes four reference evolution strategies:
adaptive_evolve,adaptive_skill,skillforge, andguided_synth. - Extensive Benchmark Support: Ready-to-use adapters for
SWE-bench,MCP-Atlas,Terminal-Bench 2.0, andSkillsBench. - Git-Integrated: Every accepted mutation is git-tagged, providing a full audit trail and easy rollback."},