Kaiyan Zhang
CTO of Frontis.AI.
Zhang Kaiyan is the Chief Technology Officer (CTO) of Frontis.AI. He received his bachelor's and master's degrees in Computer Science and Technology from Harbin Institute of Technology, and his Ph.D. from the Department of Electronic Engineering at Tsinghua University, where he studied under Professor Zhou Bowen, Director and Chief Scientist of Shanghai Artificial Intelligence Laboratory. He has received honors including the National Scholarship for Doctoral Students. His main research directions include large language models, reinforcement learning, multi-agent collaboration, and self-evolving agents. He is committed to building self-evolving agents that can continuously learn from their own experience and recursively improve themselves: researching "AI for AI" self-improvement mechanisms such as test-time reinforcement learning, multi-agent training, and reward modeling, and applying them to high-value scenarios such as enterprise and scientific discovery while building real-task evaluations. His representative works include TTRL, MARTI, NatureBench, RL4LRM, UltraMedical, and others. The open-source models and datasets he has led have been downloaded more than 400,000 times on Hugging Face, and the open-source projects he has led or participated in have cumulatively received more than 10,000 stars on GitHub. He has published more than 40 papers in top conferences and journals such as NeurIPS, ICLR, ICML, ACL, and EMNLP, with nearly 4,000 citations and an h-index of 26 on Google Scholar. He serves as the project leader of a Beijing Municipal Science and Technology Program project and as a key member of the research team for the Science and Technology Innovation 2030 "New Generation Artificial Intelligence" major project. He has long been committed to the open-source sharing and industrial application of frontier research.
Topic
Meta-Agent-Driven AI Training and Self-Evolution: AI4AI Technologies and Industry Practices
This talk focuses on the Frontis-MA series of meta-agent models and explores AI4AI-driven approaches to model training, capability expansion, and self-evolution. Covering specialized model development and Agent-based automated training, it presents technical approaches to task environment construction, autonomous experimentation, training and evaluation, and experiential learning, while examining methods for jointly optimizing model capabilities and agent systems. Using real-world tasks such as NatureBench, the talk demonstrates how to evaluate cross-task generalization and scientific problem-solving capabilities. It further draws on industry practices to analyze engineering trade-offs among effectiveness, efficiency, cost, and deployment constraints, exploring practical approaches to feeding execution experience back into model training to continuously improve agent capabilities. Outline: 1. Building Meta-Agent Capabilities with the Frontis-MA Series and AI4AI 2. Automated Development of Machine Learning Engineering Systems and Specialized Models 3. NatureBench and Real-World Scientific Problem Solving 4. Agent-Based Automated Training, Experiential Learning, and Continuous Evolution 5. Industrial R&D Delivery and Engineering Trade-Offs Audience Takeaways: - Understand intelligent production systems driven by meta-agents. - Learn practical approaches to specialized model development and Agent-based automated training. - Gain insights into connecting R&D experience, training and evaluation, and business feedback. - Develop a comprehensive evaluation perspective that balances effectiveness, efficiency, and cost.
