Title: Feedback-Guided LLM Reasoning: A Unified Closed-Loop Framework for Code Generation and Decision Making

Date: Tuesday, May 19, 2026

Time: 2:30 PM to 4:30 PM Eastern Time (US)

Location: Klaus 1315

Virtual Meeting: https://gatech.zoom.us/j/91561524771

Kamel Alrashedy

CS Ph.D. Student

School of Interactive Computing

College of Computing

Georgia Institute of Technology

https://kamel773.github.io/

Committee:

Dr. Matthew Gombolay - Advisor, School of Interactive Computing

Dr. Kartik Goyal - Georgia Tech, School of Interactive Computing

Dr. Zsolt Kira - Georgia Tech, School of Interactive Computing

Dr. Shreyes Melkote - Georgia Tech, School of Mechanical Engineering

Dr. Jesse Thomason - University of Southern California, Department of Computer Science 

Abstract:

Large Language Models (LLMs) have shown strong potential for improving productivity in software engineering, design, and decision support. However, when deployed in real-world tasks, they often struggle to produce outputs that are correct, secure, and compliant with constraints. This thesis develops Feedback-Guided LLM Reasoning (FGLR), a closed-loop interaction framework that enables language models to improve solution quality through iterative feedback from external tools, rather than relying solely on static prompting or internal model knowledge. First, I introduce CADCodeVerify, a feedback-driven framework for CAD code generation that uses vision-language verification to iteratively detect and correct structural and geometric errors. Second, I develop Feedback-Driven Security Patching (FDSP), which integrates static code analyzers with LLMs to identify vulnerabilities and iteratively refine insecure code. Third, I propose Constraints-of-Thought (Const-o-T), a structured reasoning framework that transforms natural language strategies into executable intent-constraint representations, enabling constraint-guided planning and search. Finally, I investigate reliable feedback as a fundamental challenge in feedback-driven reasoning, studying when iterative refinement improves or degrades