Keivan Rezaei

Bio

I am a third-year Ph.D. student at the University of Maryland, advised by Prof. Feizi and Prof. Hajiaghayi. My research focuses on the interpretability of generative AI models from both a model perspective—localizing knowledge within models, detecting, and explaining their failure modes—and a data perspective, analyzing the impact of individual data points on a model through challenges like unlearning and data selection for language model pretraining. Furthermore, I have proposed methods for integrating ads into the output of LLMs as a strategy to monetize them effectively.

News

[Mar 25] Our paper "Online Advertisements with LLMs: Opportunities and Challenges" is accepted to ACM SIGecom Exchanges 2025.

[Jan 25] I'll be joining Google as Student Researcher for Spring 2025 to work on data selection in large language model pretraining.

[Jan 25] My internship project "RESTOR: Knowledge Recovery through Machine Unlearning" is now on arXiv.

[Nov 24] See our blogpost on Mechanstic Interpretability Beyond Language Models!

[Sep 24] Our paper "Ad Auctions for LLMs via Retrieval Augmented Generation" is accepted to NeurIPS 2024.

[May 24] Our paper "On Mechanistic Knowledge Localization in Text-to-Image Generative Models" is accepted to ICML 2024.

[Mar 24] I will be joining MOSAIC at Allen Institute for AI (Ai2) as Research Intern in summer 2024!

Publications

arXiv

RESTOR: Knowledge Recovery through Machine Unlearning

Keivan Rezaei, Khyathi Chandu, Soheil Feizi, Yejin Choi, Faeze Brahman, Abhilasha Ravichander

arXiv (preprint)

PDF Code

ACM

Online Advertisements with LLMs: Opportunities and Challenges

^{α, β} Soheil Feizi, MohammadTaghi Hajiaghayi, Keivan Rezaei, Suho Shin

ACM SIGecom Exchanges 25

PDF

arXiv

Improving Compositional Attribute Binding in Text-to-Image Generative Models via Enhanced Text Embeddings

Arman Zarei*, Keivan Rezaei*, Samyadeep Basu, Mehrdad Saberi, Mazda Moayeri, Priyatham Kattakinda, Soheil Feizi

arXiv (preprint)

PDF

ICML

On Mechanistic Knowledge Localization in Text-to-Image Generative Models

Samyadeep Basu*, Keivan Rezaei*,Ryan Rossi, Cherry Zhao, Vlad Morariu, Varun Manjunatha, Soheil Feizi

International Conference on Learning Representations (ICML), 2024.

PDF Code

ICLR

PRIME: Prioritizing Interpretability in Failure Mode Extraction

Keivan Rezaei*, Mehrdad Saberi*, Mazda Moayeri, and Soheil Feizi

International Conference on Machine Learning (ICLR), 2024.

PDF

ICML

Text-To-Concept (and Back) via Cross-Model Alignment

Mazda Moayeri*, Keivan Rezaei*, Maziar Sanjabi, and Soheil Feizi

International Conference on Machine Learning (ICML), 2023.

PDF Code

NeurIPS

Ad Auctions for LLMs via Retrieval Augmented Generation

^{α, β} Mohammad Hajiaghayi, Sébastien Lahaie, Keivan Rezaei, Suho Shin

Neural Information Processing Systems. (NeurIPS), 2024.

PDF

ICML

Run-Off Election: Improved Provable Defense against Data Poisoning Attacks

Keivan Rezaei*, Kiarash Banihashem*, Atoosa Chegini, and Soheil Feizi

International Conference on Machine Learning (ICML), 2023.

PDF Code

ICLR

Robustness of AI-Image Detectors: Fundamental Limits and Practical Attacks

Mehrdad Saberi, Vinu Sankar Sadasivan, Keivan Rezaei, Aounon Kumar, Atoosa Chegini, Wenxiao Wang, Soheil Feizi

International Conference on Learning Representations (ICLR), 2024.

PDF

EC

Multi-agent Delegated Search

^{α, β} MohammadTaghi Hajiaghayi, Keivan Rezaei and Suho Shin

Conference on Economics and Computation (EC), 2023.

PDF

AAAI

Regret Analysis of Repeated Delegated Choice

^{α, β} MohammadTaghi Hajiaghayi, Mohammad Mahdavi, Keivan Rezaei, and Suho Shin

The Association for the Advancement of Artificial Intelligence (AAAI), 2024.

PDF