Self-improving AI Agents
Session Overview
What if AI agents could go beyond learning from a fixed set of data and continuously redesign not only how they learn, but also the systems through which that learning takes place? This session introduces EXAONE Foundry, an enterprise-grade intelligence production platform designed to enable self-improving AI agents for domain experts and professional environments.
EXAONE Foundry connects the full intelligence-development lifecycle into a continuous feedback loop. This includes generating and evaluating domain-specific training data, automatically exploring optimal hyperparameters and learning strategies, designing system architectures around expert workflows, and assessing performance through evaluation criteria and metrics derived from expert behavior and judgment.
The platform is designed to bridge the gap between tacit expertise—knowledge that specialists possess but may struggle to articulate—and knowledge that can be explicitly represented and operationalized. It also addresses the practical limitation that no individual expert can exhaustively explore every possible combination of data, strategy, model, tool, and system architecture.
Rather than improving only the model, EXAONE Foundry repeatedly refines the entire environment surrounding the agent, including data, memory, tools, orchestration, evaluation frameworks, and learning systems. This allows AI agents to identify their own limitations, generate the experiences and tasks required to overcome them, discover signals that are difficult for humans to detect, and evolve continuously alongside human expertise.
Key Takeaways
- How AI agents can move beyond static learning and continuously redesign their own learning processes
- How EXAONE Foundry connects data generation, training strategy, expert workflows, and evaluation into a continuous intelligence-development loop
- How tacit expert knowledge can be translated into measurable criteria, tools, and system behavior
- Why self-improving agents require the continuous optimization of data, memory, tools, orchestration, evaluation, and the learning environment—not just the model itself
- How competitive and collaborative domain-specific reinforcement learning environments may help provide a path from AGI toward superintelligence
Speaker
Moontae Lee is Vice President and Head of the Superintelligence Lab at LG AI Research. His work focuses on developing advanced AI systems that can continuously improve their learning strategies, tools, evaluation methods, and system architectures.
Before joining LG AI Research, he was a Research Scientist at Microsoft Research Redmond and an Assistant Professor at the University of Illinois Chicago.
He holds a master’s degree from Stanford University and a Ph.D. in Computer Science from Cornell University.