Post by National AI Research Lab (NAIRL)

How should an AI balance goals that pull against each other, and how does it learn when it can only work from a fixed dataset? Researchers from Professor Youngchul Sung's team at KAIST, affiliated with the National AI Research Lab (NAIRL), are presenting two papers at #ICML2026 and #ICLR2026 on these questions in reinforcement learning. The first study, "Constrained Multi-Objective Reinforcement Learning with Max-Min Criterion" (#ICML2026), deals with AI that has to serve several conflicting goals at once. A common way to keep things fair is to focus on improving whichever goal is doing worst, but that has been hard to combine with hard constraints the system cannot break. The team proposes a framework that does both, and shows it works across building thermal control, locomotion, and greenhouse-gas-aware traffic management. The second study, "Flow Actor-Critic for Offline Reinforcement Learning" (#ICLR2026), looks at learning from a fixed dataset with no further interaction. These datasets are often too complex for simple policies to capture, so the team built a method using a more expressive flow model for both the actor and the critic. It reaches state-of-the-art results on offline RL benchmarks including D4RL and OGBench. We extend our appreciation to Professor Sung and the research team for their contribution to this important area of AI research. 🔗 Paper (ICML 2026): arXiv link forthcoming 🔗 Paper (ICLR 2026): https://lnkd.in/g8EFwKm3