Can AI Agents Learn from Their Own Failures? Toward Recursive Self-Improvement by Dr Xiaoxiao Li

Abstract

Agents fail in ways that are hard to debug: a tool call goes wrong, a step gets skipped, and the whole chain breaks three stages later. Today the fix is usually a human reading traces and rewriting prompts by hand. This talk is about getting agents to do that work themselves. I will show how to trace a failure back to the step that cause it, how agents can write and improve their harness, and introduce our evaluation platform that turns each failure into a recipe for fixing it.

About the Speaker

Dr. Xiaoxiao Li is an Associate Professor in the Department of Electrical and Computer Engineering at the University of British Columbia, a Faculty Member at the Vector Institute, and a Visiting Faculty Member at Google. Dr. Li holds a Canada Research Chair (Tier II) in Responsible AI and is recognized as a Canada CIFAR AI Chair. Dr. Li's current interests include mechanistic analysis of modern deep learning methods, developing hypothesis-driven evaluations, and advancing methodologies toward artificial general intelligence.