Continual Learning for AI Agents: From Failures to Durable Improvements - Soheil Feizi, RELAI
This talk introduces verifiable continual learning (VCL) for AI agents, a method to achieve durable improvements from experience without forgetting past performance. It addresses the challenges of obtaining feedback from production logs and acting upon it effectively. VCL aims to transform failures into testable scenarios, ensuring that improvements are verified and do not introduce regressions.
World's Fair 2026 23 min