Ph.D. Thesis Proposal

 

Bo Yuan

(Faculty advisor: Professor Yongxin Chen) 

 

"Generative LLM Agent Systems: From Workflows to Self-Evolving Agents"

 

Wednesday, April 29

2:00 - 4:00 p.m.

Coda C1015

Virtual

https://gatech.zoom.us/j/95995309752?pwd=XPsKUrHaX1MdYMsnbGv9A8tRRqiP7I.1&from=addon

 

Abstract:

Large language models serve as the foundation of complex agent systems that coordinate model calls, external tools, memory, and domain-specific modules to solve tasks beyond the capabilities of any individual model. This thesis investigates generative LLM agent and traces a progression from carefully designed inference-time workflows to self-evolving agents that improve their own policies, tool libraries, and training data over time. The first part of the thesis studies automatic workflow generation and multi-agent alignment by formulating them as problems of posterior sampling and evidence-based posterior contraction, respectively, yielding information-theoretic guarantees on coordination reliability. The second part develops a proactive two-agent text-to-CAD framework that identifies ambiguity in underspecified prompts, elicits targeted clarifications, and synthesizes executable programs. Building on these foundations, the proposed research aims to close the learning-time loop over both policy and data, enabling self-evolving agents that iteratively refine their capabilities and are evaluated on frontier scientific discovery benchmarks, as well as through an autoresearch agent for autonomous dataset creation. 

Committee:

• Dr. Yongxin Chen (Advisor), School of Aerospace Engineering, Georgia Institute of Technology
• Dr. Kartik Goyal, School of Interactive Computing, Georgia Institute of Technology
• Dr. Wei Xu, School of Interactive Computing, Georgia Institute of Technology