Training Agents IRL

GitHub and Hugging Face are hosting an in-person Open Source AI Week session about how AI agents are trained and evaluated. The discussion goes beyond one-shot reinforcement learning to multi-step, environment-driven training, covering reward design, rollouts, benchmarks and the gap between laboratory gains and agents' real-world behavior.

The evening consists of short talks with questions followed by a panel on what currently works in reinforcement learning for agents and where the bottlenecks remain. It is particularly relevant for ML practitioners, agent builders and reinforcement-learning researchers. Registration is approval-required, and attendees should use the official event page to confirm access.