Date: November 24th, 2026

Time: 11:00 AM – 1:00 PM ET

Location: Zoom Meeting

Siheng Xiong

Machine Learning PhD Student

School of Electrical and Computer Engineering

Georgia Institute of Technology

Committee

1. Dr. Faramarz Fekri (Advisor), Georgia Institute of Technology
2. Dr. Ghassan AlRegib, Georgia Institute of Technology
3. Dr. Kartik Goyal, Georgia Institute of Technology
4. Dr. Suguman Bansal, Georgia Institute of Technology
5. Dr. Karthikeyan Sundaresan, Georgia Institute of Technology

Abstract

Large language models (LLMs) have demonstrated strong capabilities in complex problem-solving but still face challenges in reliable multi-step reasoning, planning, and self-improvement. While chain-of-thought (CoT) prompting enables intermediate reasoning, it often lacks explicit structures to guide and verify the reasoning process. We introduce structure-driven agentic reasoning, a paradigm that integrates structured representations with adaptive reasoning and learning to improve LLM capabilities. Specifically, we explore graph-based reasoning for temporal understanding and logical verification, hierarchical planning and execution, adaptive information acquisition through reinforcement learning, and experience-driven skill evolution. Extensive experiments across temporal, mathematical, logical, and knowledge-intensive tasks demonstrate improvements in reasoning accuracy, efficiency, and skill transfer. Together, these findings highlight the potential of explicit structure to support more reliable, efficient, and adaptive reasoning in LLMs.