Kazem Faghih
PhD Student in Computer Science
University of Maryland, College Park
Advised by Prof. Soheil Feizi
I'm a Computer Science PhD student at the University of Maryland, College Park, advised by Prof. Soheil Feizi. I work on LLM evaluation, reliability, and agentic AI. I study how model behavior changes across equivalent user requests, tool descriptions, and agentic interactions, and I develop better ways to evaluate and improve the reliability of these systems.
Before my PhD, I worked across machine learning research, software engineering, and large-scale data systems, including GPU-accelerated sequence search at the Max Planck Institute and industry work on language models, retrieval-augmented generation, and distributed systems.
Open to Summer 2027 research internships in LLM research, AI evaluation, reliability, and agentic AI.
- kazemf@umd.edu
- Scholar
- GitHub
- CV (soon)
Research
LLM evaluation & reliability
Looking beyond benchmark accuracy to ask whether a model behaves consistently when equivalent requests are phrased differently.
Agentic & tool-augmented LLMs
How agents choose and use tools, how sensitive those choices are to tool descriptions and context, and the security of the ecosystems agents rely on.
Better evaluation methods
Developing evaluation methods that show where a model can be relied on and where it falls short, so that reliability can be measured and improved.
News
- 2026Your LLM Agents are Temporally Blind accepted to Findings of ACL 2026.
- 2025Tool Preferences in Agentic LLMs are Unreliable accepted to the EMNLP 2025 Main Conference.
- Jan 2025Started my PhD at the University of Maryland with Prof. Soheil Feizi, supported by a Dean’s Fellowship.
Publications
- arXiv
- arXiv
- Findings of ACL 2026
- arXiv
- EMNLP 2025
Tool Preferences in Agentic LLMs are Unreliable
Shows that surface-level edits to tool descriptions can raise selection probability by more than 20× in some settings.
Experience
University of MarylandResearch
Research Assistant · Advisor: Prof. Soheil Feizi
Research on LLM and agent reliability: how model behavior varies across user formulations, tool interfaces, and agentic settings.
YektanetIndustry
Software Engineer
Built retention-marketing infrastructure for 100M+ contacts delivering 100K+ SMS/push messages per minute, and a user-location service at ~4K requests/second.
Mofid SecuritiesIndustry
ML Engineer & Data Scientist
Led development of a Persian generative language model (data curation, pre-training, fine-tuning, evaluation) and built a production retrieval-augmented chatbot.
Max Planck Institute for Multidisciplinary SciencesResearch
Research Assistant · Prof. Johannes Söding, with Prof. Martin Steinegger
Developed a new GPU prefilter for MMseqs2, the protein sequence-search tool used in workflows such as AlphaFold and ColabFold, using CUDA-parallelized ungapped alignment. It ran about 2× faster than the optimized CPU version at comparable sensitivity.
Sharif University of TechnologyResearch
Research Assistant · Prof. Ali Sharif-Zarchi
Studied representation learning for peptide-HLA binding prediction using protein language models, AlphaFold-derived representations, and multiple-sequence alignments.
Cafebazaar / Balad MapsIndustry
Software Engineer & Data Engineer
Ran backend and data services processing ~200M map records per day; migrated high-throughput services from Django/REST to Go/gRPC on Kubernetes.
NimboIndustry
Big Data Intern
Built a distributed search-engine prototype over ~100M documents with Hadoop, Spark, Kafka, Elasticsearch, and PageRank.
Education
University of Maryland, College Park
Ph.D. in Computer Science
Advisor: Prof. Soheil Feizi · Dean’s Fellowship
Sharif University of Technology
B.Sc. in Computer Engineering
Tehran, Iran
Honors & Service
- 2025Dean’s Fellowship, University of Maryland. Awarded on admission for outstanding academic achievement.
- 20172nd place nationally out of 145,000+ participants (Mathematics & Physics) in Iran’s university entrance exam (Konkour).
- ServiceReviewer: ACL, NeurIPS