Reza Fayyazi

Reza Fayyazi

PhD Candidate · Electrical & Computer Engineering · Rochester Institute of Technology · Rochester, NY

Email Scholar GitHub LinkedIn

About

I am a PhD Candidate in Electrical and Computer Engineering at Rochester Institute of Technology (RIT), working at the intersection of Large Language Models (LLMs), Mathematics, and Trustworthy AI.

My research focuses on making the reasoning of LLMs auditable by developing explainable techniques to verify provenance, trace reasoning processes, and attribute AI-generated responses to their underlying sources.

My broader goal is to improve decision-making and build trust in Agentic-AI applications by making how they derive their responses more transparent and mathematically traceable.

Research Interests

Explainable & Trustworthy AI

Methods that make a model's behavior traceable rather than asking users to trust it.

Agentic-AI Systems

LLM agents that plan and act on complex tasks under structured, auditable constraints.

LLM Reasoning & Alignment

How to adaptively align LLM behavior with human intent and values.

Adaptability in High-Stakes Environments

Systems that remain reliable when there are security issues and unexpected changes.

Topological Data Analysis

Understanding the shape and structure of data to reveal hidden patterns and relationships.

Deep Learning

Neural network architectures and training methods for advanced AI applications.

Publications

2026 Topological Attribution Distance (TAD): Revealing Segment-Level RAG Influence on LLM Output Geometry for Incident Log Analysis Fayyazi, R., Zuzak, M., Yang, S. J. arXiv· under review Powers TAD 2026 How Can You Tell if Your Large Language Model Could Be a Closet Antisemite? An Explainability-Based Audit Framework for Implicit Bias Dutta, A., Fayyazi, R., Yang, S., KhudaBukhsh, A. R. AAAI-2026 · Special Track on AI for Social Impact 2025 LLM Embedding-based Attribution (LEA): Quantifying Source Contributions to Generative Model's Response for Vulnerability Analysis Fayyazi, R., Zuzak, M., Yang, S. J. arXiv · pending journal review Powers LEA 2025 ProveRAG: Provenance-Driven Vulnerability Analysis with Automated Retrieval-Augmented LLMs Fayyazi, R., Hoyos Trueba, S., Zuzak, M., Yang, S. J. IEEE Access Powers ProveRAG 2025 Guided Reasoning in LLM-Driven Penetration Testing Using Structured Attack Trees Nakano, K., Fayyazi, R., Yang, S. J., Zuzak, M. COLM 2025 2024 Advancing TTP Analysis: Harnessing the Power of Large Language Models with Retrieval Augmented Generation Fayyazi, R., Taghdimi, R., Yang, S. J. ACSAC Workshops 2024 Is More Information Better? How Increasingly Augmented LLM Help in Vulnerability Assessment Fayyazi, R., Yang, S. J. USENIX Security · poster 2023 On the Uses of Large Language Models to Interpret Ambiguous Cyberattack Descriptions Fayyazi, R., Yang, S. J. arXiv 2022 Translating Cybersecurity Descriptions into Interpretable MITRE Tactics using Transfer Learning Fayyazi, R., Wufeng, Y. S., Gautam, P., Yang, S. J. ACSAC 2022 · poster

Full list on Google Scholar →

Introducing AuditLM

To be able to verify AI reasonings and developing trust in Agentic-AI applications, we combined our research publications, namely ProveRAG, LEA, and TAD, into a single framework called AuditLM.

AuditLM is a framework for auditing LLM-generated responses, and it is designed to be modular and extensible. It provides three auditing lenses: ProveRAG for provenance verification, LEA for token-level source attribution, and TAD for segment-level evidence verification.

How It Works

Visit AuditLM