• [11/29/2025] The code of Agent0 was released!
• [11/26/2025] We’ve set up a Discord server and WeChat group to make it easier to collaborate and exchange ideas on this project. Welcome to join the Group to share your thoughts, ask questions, or contribute your ideas! 🔥 Join our Discord and WeChat Group Now!
• [11/25/2025] Agent0-VL was released on arXiv!
• [11/20/2025] Agent0 paper was released on arXiv!

The Agent0 Series explores a new direction for autonomous agent development, showing that capable agents can improve and evolve without relying on human-curated datasets or handcrafted supervision. This repository brings together two complementary studies that advance self-improving agents through tool-integrated reasoning.

Unleashing Self-Evolving Agents from Zero Data via Tool-Integrated Reasoning

A fully autonomous framework that evolves high-performing language agents through multi-step co-evolution and seamless tool integration. Agent0 establishes a symbiotic competition between two agents:

• Curriculum Agent: Proposes increasingly challenging frontier tasks
• Executor Agent: Learns to solve them using external tools

Key Results:

• ✅ +18% improvement on mathematical reasoning benchmarks
• ✅ +24% improvement on general reasoning benchmarks
• ✅ Zero external data required for training
• ✅ Multi-turn interaction support

📄 Paper | 📁 Code | 🔗 Details

Exploring Self-Evolving Agent for Tool-Integrated Vision-Language Reasoning

A self-evolving vision-language agent that extends the Agent0 paradigm to multimodal reasoning tasks. Agent0-VL incorporates tool usage not only into reasoning but also into self-evaluation and self-repair through a dual-role architecture:

• Solver: Performs multi-turn tool-integrated reasoning
• Verifier: Generates structured feedback and fine-grained self-rewards

Key Results:

• ✅ +12.5% average improvement on visual reasoning benchmarks
• ✅ +7.3% improvement in test-time scaling performance
• ✅ State-of-the-art among open-source vision-language models
• ✅ Zero external reward for self-evolution

📄 Paper | 📁 Code | 🔗 Details

Both Agent0 and Agent0-VL are built on the principle of zero-data self-evolution:

• No Human Annotations: Completely eliminates dependency on external data or human supervision
• Tool-Integrated Reasoning: Leverages external tools to enhance problem-solving capabilities
• Autonomous Evolution: Self-generates training data through intelligent exploration

Complete comparison with state-of-the-art self-evolving methods:

Key Improvements:

• 📈 +18.3% over base model (49.2 → 58.2)
• 🎯 +6.4% over R-Zero (54.7 → 58.2)
• 🔥 +3.7% over Socratic-Zero (56.1 → 58.2)

Key Improvements:

• 📈 +22.0% over base model (34.5 → 42.1)
• 🎯 +5.5% over Absolute Zero (39.9 → 42.1)
• 🔥 Highest overall performance among all self-evolving methods

Comprehensive comparison with closed-source and open-source models:

Key Improvements (Agent0-VL-7B):

• 📈 +12.5% over Qwen2.5-VL-7B base (58.3 → 65.6)
• 🎯 +10.3% over Qwen2.5-VL-7B-TIR (59.5 → 65.6)
• 🔥 +4.3% over ThinkLite-VL-7B (62.9 → 65.6)
• 🏆 Best among all open-source 7B models

Key Improvements (Agent0-VL-8B):

• 📈 +6.1% over Qwen3-VL-8B base (70.3 → 74.6)
• 🎯 +4.6% over Qwen3-VL-8B-TIR (71.3 → 74.6)
• 🔥 Outperforms GPT-4o on MathVista, HallBench, and ChartQA
• 🏆 State-of-the-art among all open-source models

Evolution Progress:

• 🔄 Iter 1: +5.2% improvement (57.3 → 60.5)
• 🔄 Iter 2: +4.0% additional gain (60.5 → 63.6)
• 🔄 Iter 3: +2.8% further improvement (63.6 → 65.5)
• ✅ +8.2% cumulative gain over base model

If you find our work helpful, please consider citing:

@article{xia2025agent0,
  title={Agent0: Unleashing Self-Evolving Agents from Zero Data via Tool-Integrated Reasoning},
  author={Xia, Peng and Zeng, Kaide and Liu, Jiaqi and Qin, Can and Wu, Fang and Zhou, Yiyang and Xiong, Caiming and Yao, Huaxiu},
  journal={arXiv preprint arXiv:2511.16043},
  year={2025}
}

@article{liu2025agent0vl,
  title={Agent0-VL: Exploring Self-Evolving Agent for Tool-Integrated Vision-Language Reasoning},
  author={Liu, Jiaqi and Xiong, Kaiwen and Xia, Peng and Zhou, Yiyang and Ji, Haonian and Feng, Lu and Han, Siwei and Ding, Mingyu and Yao, Huaxiu},
  journal={arXiv preprint arXiv:2511.19900},
  year={2025}
}

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

We thank the open-source community for their foundational work that made this research possible. Special thanks to:

• The teams behind Qwen, InternVL, and other base models
• The VeRL team for their excellent RL framework
• All the benchmark creators and maintainers