A-EVO-Lab/a-evolve

The official repository of "Position: Agentic Evolution is the Path to Evolving LLMs".

What it solves

A-Evolve 提供了一套用於 AI Agent 自主自我改進的通用基礎設施。它消除了對手動搭建測試框架(harness engineering)的需求,讓任何 Agent 可以透過各種演化演算法,在不同領域中演化其自身的提示詞、技能與記憶。

How it works

該系統將 Agent 可演化的狀態——例如系統提示詞、技能庫與情節記憶(episodic memory)——視為檔案系統上的標準目錄結構(即「Agent Workspace」)。接著,演化引擎會執行一個持續的迴圈:它讓 Agent 解決任務,觀察結果,分析效能日誌以突變工作空間檔案,並在透過 git-tagging 提交之前,於保留任務(holdout tasks)上驗證這些突變。

Who it’s for

它專為從事 LLM 自我優化、強化學習與 Agent 架構研究的 AI 研究人員與開發者設計,旨在讓他們在無需手動調優的情況下,測試與演化 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 the loop.
  • Built-in Algorithms: Includes four reference evolution strategies: adaptive_evolve, adaptive_skill, skillforge, and guided_synth.
  • Extensive Benchmark Support: Ready-to-use adapters for SWE-bench, MCP-Atlas, Terminal-Bench 2.0, and SkillsBench.
  • Git-Integrated: Every accepted mutation is git-tagged, providing a full audit trail and easy rollback."},