AgentOdyssey：解鎖大模型持續學習的開放式文字遊戲生成引擎 | AgentOdyssey: An Open-Ended Text Game Engine for Test-Time Continual Learning Agents
🔎 工具速覽 / AT A GLANCE
| Category | AI Agent Evaluation Framework / LLM Research Tool |
| Pricing | Free / Open Source |
| BestFor | AI researchers studying continual learning, long-horizon planning, and autonomous agents. |
| GitHub Stars | ⭐ 37 |
🚀 引言 / Introduction
AgentOdyssey 是一個輕量級的交互式環境，旨在挑戰 AI Agent 在未知環境中的適應能力。它不僅能生成全新的長程文字遊戲，更提供了一套嚴謹的評估機制，用以衡量 Agent 在測試時的持續學習表現。 / AgentOdyssey is a lightweight interactive environment designed to challenge the adaptability of AI agents in unknown environments. It not only generates novel long-horizon text games but also provides a rigorous evaluation mechanism to measure the test-time continual learning performance of agents.
🛠️ 核心功能 / Key Features
Open-Ended Long-Horizon Game Generation: Create rich games with novel entities, dynamics, and storylines via a single command.開放式長程遊戲生成：僅需單一指令即可生成具有全新實體、動態機制與豐富劇情的遊戲。
Unified Agent Interface: Standardized framework using inherited classes to ensure fair comparisons across different LLM-based agents.統一 Agent 接口：透過繼承類別實現標準化框架，確保不同 LLM Agent 之間在提示詞共享與對比時的公平性。
Multifaceted Evaluation Metrics: Advanced probing of failure modes beyond simple game progress, focusing on cognitive abilities.多維度評估指標：超越單純的遊戲進度，提供深層指標以探測 Agent 的特定失效模式與認知能力。
💡 技術亮點 / Tech Highlights
Focuses on five key abilities: exploration, world knowledge acquisition, episodic memory, skill learning, and long-horizon planning.聚焦五大核心能力：探索能力、世界知識獲取、情節記憶、技能學習以及長程規劃。
Highly extensible architecture allowing researchers to implement new agents by simply overriding a few methods.高擴展性架構，研究人員僅需實現少量方法即可快速部署並測試新的 Agent 方案。
📦 快速上手 / Quick Start
Install: Clone the repository and install via 'pip install -e .'安裝：克隆儲存庫並使用 'pip install -e .' 進行安裝
Configuration: Set your LLM API key (e.g., OPENAI_API_KEY) in the environment.配置：在環境變數中設置 LLM API 金鑰（例如 OPENAI_API_KEY）
Execution: Use 'python eval.py' to run evaluations or manually play the game with 'HumanAgent'.執行：使用 'python eval.py' 運行評估，或透過 'HumanAgent' 親自體驗遊戲
準備好試試 AgentOdyssey：解鎖大模型持續學習的開放式文字遊戲生成引擎 | AgentOdyssey: An Open-Ended Text Game Engine for Test-Time Continual Learning Agents 了嗎？
Ready to try AgentOdyssey：解鎖大模型持續學習的開放式文字遊戲生成引擎 | AgentOdyssey: An Open-Ended Text Game Engine for Test-Time Continual Learning Agents?
前往 GitHub 頁面 →
留言
張貼留言