Reza (Rey) Sanayei

Hello! My name is Rey, and I am a Computer Science graduate student at the University of Arizona. My research centers on reasoning, faithfulness, and post-training methods in large language models (LLMs). I work in the CLU Lab with Dr. Mihai Surdeanu and Dr. Ellen Riloff, where I explore how LLMs can be guided to reason more faithfully and align better with human expectations.

Previously, I collaborated with Dr. Steven Bethard on projects tackling hallucination detection and distinguishing between machine-generated and human-generated text, which strengthened my interest in evaluating and improving the reliability of AI systems.

I also bring professional experience from machine learning and software engineering internships, where I worked on applied problems ranging from traffic insights analysis to backend service optimization.

Beyond research, I served as President of the Google Developer Student Club at my university, building a vibrant community around technology and innovation. I have also been a College of Science Ambassador and Peer Mentor, mentoring students and helping them thrive academically.

Outside of work, I’m a soccer player and fan, enjoy action-adventure video games, love exploring new places, and have a music taste that spans from rap to classical.

news

Sep 22, 2025	Excited to share that my paper on non-linear reasoning with LLMs was accepted to EMNLP 2025 Findings! 🎉

selected publications

Can LLMs Judge Debates? Evaluating Non-Linear Reasoning via Argumentation Theory Semantics

Reza Sanayei, Srdjan Vesic, Eduardo Blanco, and 1 more author

In Findings of the Association for Computational Linguistics: NAACL 2025, Nov 2025

Abs arXiv

Large Language Models (LLMs) excel at linear reasoning tasks but remain underexplored on non-linear structures such as those found in natural debates, which are best expressed as argument graphs. We evaluate whether LLMs can approximate structured reasoning from Computational Argumentation Theory (CAT). Specifically, we use Quantitative Argumentation Debate (QuAD) semantics, which assigns acceptability scores to arguments based on their attack and support relations. Given only dialogue-formatted debates from two NoDE datasets, models are prompted to rank arguments without access to the underlying graph. We test several LLMs under advanced instruction strategies, including Chain-of-Thought and In-Context Learning. While models show moderate alignment with QuAD rankings, performance degrades with longer inputs or disrupted discourse flow. Advanced prompting helps mitigate these effects by reducing biases related to argument length and position. Our findings highlight both the promise and limitations of LLMs in modeling formal argumentation semantics and motivate future work on graph-aware reasoning.