On November 28, the 48th AIR Academic Salon, co-hosted by Tsinghua University's Institute for AI Industry Research (AIR) and AAR Corp, successfully took place. Professor Liu Yang, AAR Corp Chair Professor at Tsinghua University, Executive Dean of AIR, and Co-Executive Dean of the AI Hospital, delivered a keynote titled "Large Model-Driven Evolvable Agents." He discussed the development trends of large models and intelligent agents, systematically presenting his team's latest research on evolvable agents and practical explorations like the AI Hospital, while delving into questions such as "Can agents evolve continuously in real-world environments like humans?"
Professor Liu Yang is a leading figure in AI and natural language processing, with accolades including the National Science Fund for Distinguished Young Scholars. His research spans AI for Science and major national projects, earning him multiple awards and prestigious academic roles.
The rapid advancement of large language models has ushered AI into a new era. Large models serve as the "soul" of intelligence, while agents act as the "carrier" for real-world applications. In complex, dynamic environments, autonomous agents may evolve like humans, potentially surpassing human adaptability. Professor Liu proposed a framework for agent evolution based on alignment with human intent, environmental laws, and resource constraints, covering three dimensions: IQ (task competence), EQ (social interaction), and organizational evolution (multi-agent collaboration).
For IQ evolution, the team developed methods for error-driven learning and self-annotated data generation, enabling agents to learn from mistakes and improve continuously. In EQ evolution, agents acquired strategic interaction skills through language-based games like poker and werewolf, demonstrating emergent social behaviors. Organizational evolution focused on forming elite multi-agent teams that outperform individuals in complex tasks.
A practical application is the AI Hospital, where virtual doctors evolve via these mechanisms, showing measurable performance gains. This approach, inspired by reinforcement learning and multi-agent systems, accelerates medical training and optimizes healthcare workflows.
Looking ahead, Professor Liu envisioned a "second emergence of intelligence" through large-scale agent collaboration, potentially solving grand challenges beyond individual model capabilities. AIR continues to pioneer AI-driven industrial innovation under the leadership of academician Zhang Ya-Qin, focusing on smart transportation, IoT, and healthcare.
Comments