Thanks for posting our work Rohan Paul!

🚨 New Work: Agent0 - Can LLM Agents evolve from scratch with zero human data? We introduce a framework that breaks the "knowledge ceiling" of agent self-evolving by integrating tools into the loop. Core Innovations: 1️⃣ Co-Evolution: A Curriculum Agent proposes frontier tasks,

This Stanford University paper just broke my brain. They just built an AI agent framework that evolves from zero data no human labels, no curated tasks, no demonstrations and it somehow gets better than every existing self-play method. It’s called Agent0: Unleashing

Agent0 lets two LLM agents co-evolve by creating and solving harder tasks, boosting reasoning skills without any external data. Paper👉 huggingface.co/papers/2511.16…

Thanks Robert Youssef for sharing our work! Nice summary!

Heading to San Diego for #NeurIPS2025! I'll be there Dec 2–6. 🌴 Extremely interested in multimodal agents, tool-use, self-evolving frameworks and deep research. Let's talk if u are interested! ☕️ P.S. Searching for 2026 Summer Internships. Let's connect! 🤝

Multimodal version of Agent0 has also been released Jiaqi Liu. Paper: arxiv.org/abs/2511.19900 Code: github.com/aiming-lab/Age…

斯坦福大学的这篇论文值得了解👇🏻 他们构建了一个AI智能体框架，从零数据起步，没有人工标注，没有精心设计的任务，也没有任何演示，但它竟然超越了所有现有的自博弈方法。 这个框架名为Agent0：通过工具集成推理，从零数据中释放自我进化的智能体。 它所取得的成就令人难以置信。

Agent0-VL is a self-evolving agent for tool-integrated visual reasoning. This novel framework achieves continuous improvement in multimodal tasks with zero human supervision, leveraging a Solver-Verifier cycle for self-evaluation and repair.

Heading to San Diego for #NeurIPS2025! I'll be there Dec 2–6. 🌴 Extremely interested in multimodal agents, tool-use, self-evolving frameworks and deep research. Let's talk if u are interested! 📷 P.S. Searching for 2026 Summer Internships. Let's connect! 📷🤝

🚀Just released! What if a VLM could audit and improve itself — with zero human labels? Our new work Agent0-VL shows that multimodal agents can self-evolve by using tools not only to solve problems, but also to verify and repair. Paper: arxiv.org/abs/2511.19900

A great paper. Shows that a vision language model can reliably train and improve itself using tools as its own teacher, without human feedback. On key visual math benchmarks it gains about 12.5% accuracy over the starting model. The issue is that normal vision language models

🎉 We already released our code in github.com/aiming-lab/Age…. If you have any questions, please feel free to ask or submit an issue.

8B model can outperform AlphaEvolve on open optimization problems by scaling compute for inference or test-time RL🚀! ⭕Circle packing: AlphaEvolve (Gemini-2.0-Flash/Pro) : 2.63586276 Ours (DeepSeek-R1-0528-Qwen3-8B) : 2.63598308 🔗in🧵 [1/n]

Peng (Richard) Xia ✈️ NeurIPS 25 Huaxiu Yao ✈️ NeurIPS 2025 Jiaqi Liu ✈️ NeurIPS 2025 Yiyang Zhou Can Qin Caiming Xiong Fang Wu Another week, another paper claiming we finally cracked autonomous AI. Agent0: two copies of the same base model play an elaborate game of curricular ping-pong inside their context window, using a Python REPL as the ball. No external data, no human labels, no lasting updates