Tracing the Path to Self-adapting AI Agents

Watch a 57-minute colloquium presentation where Senior Researcher Ching-An Cheng from MSR AI Frontiers introduces Trace, an innovative AutoDiff-like framework for training AI workflows end-to-end. Learn how Trace extends beyond traditional back-propagation by capturing and propagating AI workflow execution traces while leveraging LLM-based optimization to enhance performance. Discover how this PyTorch-like Python library seamlessly integrates with any Python workflow, allowing developers to optimize specific components such as code and prompts similar to neural network training. Gain insights into the framework's design principles and see practical demonstrations of Trace's capabilities in training AI agents. The speaker, who holds a PhD in Robotics from Georgia Tech, brings extensive expertise in sequential decision-making, robotics, and developing foundations for practical algorithms that address real-world challenges. His recent work focuses on enabling agents to learn from general feedback, unifying concepts from Language Feedback Learning, reinforcement learning, and imitation learning, with his research garnering prestigious awards including the ICML 2022 Outstanding Paper Award Runner-Up and AISTATS 2018 Best Paper.