FN.

PhD student · The Ohio State University

Fatemeh NourzadI also go by Atena.

I am a Ph.D. student in Electrical and Computer Engineering at The Ohio State University, advised by Professor Ness B. Shroff. I study how AI systems can acquire new capabilities and adapt to different goals while preserving their reasoning and reliable behavior.

My interest in engineering began with childhood conversations with my father, an earthquake-engineering professor. I found my own path through electrical engineering, drawn to the probability and optimization behind communication networks. Reading about machine learning for network optimization led me toward ML itself.

During my master’s at the University of Tehran, I worked with the International Institute of Earthquake Engineering and Seismology to detect earthquakes in noisy recordings. That experience sparked an interest in reliability that I carried into safe reinforcement learning at Ohio State, then into language-model alignment.

In collaboration with researchers at Google, I studied how one language model can adapt to different human preferences without retraining. Through AI-EDGE, I also worked on distributed LLM training, bringing together my interests in learning algorithms and communication networks.

More recently, I have investigated how learning to use tools changes LLM agents’ reasoning, and developed methods to help them retain these abilities. I am now exploring recursive self-improvement in multi-agent systems, asking how agents can improve collectively while their underlying models remain fixed.

Research

Tool-using agents & reasoning retention

How does learning to use tools change an agent’s reasoning? We investigated how tool-use reinforcement learning reshapes reasoning, finding that it can shorten deliberation even when its effects on accuracy vary across tasks. We developed methods to help agents acquire tool-use skills while retaining their ability to reason independently.

Preference-conditioned & federated alignment

We study how language models can balance competing objectives, from adapting to individual preferences to learning collaboratively across clients.

Meta-learning for preference steering

In collaboration with researchers at Google, we developed a meta-learning framework that trains a single language-model policy to respond to different preference combinations at inference time, without retraining. We evaluated how well it balances competing objectives and generalizes to unseen preferences, and investigated optimization and geometric barriers that can limit continuous steering.

Federated multi-objective alignment

How can clients align a shared language model when their objectives conflict and their data remain local? We developed a federated multi-objective alignment framework in which clients solve regularized optimization problems locally. This controls cross-client objective disagreement and reduces communication by avoiding the transmission of separate gradients for every objective. We established finite-time Pareto-stationarity bounds and evaluated the method across heterogeneous clients.

Learning across distributed systems

Institutions have useful compute, but rarely share identical hardware or networks. Through AI-EDGE, I co-developed systems for pretraining and federated fine-tuning across geographically distributed GPU clusters.

Earlier research

Master’s Research · Seismic Event Detection & Classification

University of Tehran · Computational Audio-Vision Lab

Working with the International Institute of Earthquake Engineering and Seismology, I developed a deep-learning pipeline to detect seismic events, distinguish earthquakes from noise, and extract event features.

Undergraduate Research · Learning-Based Smart Thermostat

University of Tehran · Multimedia Communication Lab

Built and evaluated a Raspberry Pi–based thermostat that uses neural networks to learn occupants’ habits and adapt temperature settings for comfort and energy efficiency.

Papers & manuscripts

Publications.

*Co-primary authors with equal contribution in alphabetical order.

Thinking Beyond Tools: How Tool-Use RL Reshapes Deliberation

Fatemeh Nourzad*, A. Roknilamouki*, Q. Wu*, E. Ekici, Y. Liang, N. B. Shroff

Manuscript under review · 2026

Related workshop contributions

  • Selective Forgetting in Tool-Augmented Language-Model Agents

    Fatemeh Nourzad*, A. Roknilamouki*, Q. Wu*, E. Ekici, Y. Liang, N. B. Shroff

    NeurIPS 2026 TTCL Workshop
    CoLLAs 2026 Work-In-Progress Track

    Paper ↗

Learning to Steer the Pareto Frontier: Meta-Learning for Multi-Objective LLM Alignment

Fatemeh Nourzad, D. Sow, Y. Liang, M. Zhang, Y. Li, E. Ekici, N. B. Shroff

Manuscript under review · 2026

Related workshop contributions

  • Barriers to Pareto Steerability in Preference-Conditioned LLM Alignment

    Fatemeh Nourzad, D. Sow, Y. Liang, M. Shi, M. Zhang, Y. Li, E. Ekici, N. B. Shroff

    ICLR 2026 I Can’t Believe It’s Not Better Workshop

    Paper ↗
  • One Model, Many Goals: Meta-Learning Preference-Conditioned Alignment for Lifelong LLM Agents

    Fatemeh Nourzad, D. Sow, Y. Liang, M. Shi, M. Zhang, Y. Li, E. Ekici, N. B. Shroff

    ICLR 2026 Lifelong Agents Workshop

    Paper ↗

Rethinking Multi-objective Resolution in Federated LLM Alignment

Fatemeh Nourzad, A. Roknilamouki, H. Yang, E. Ekici, J. Liu, N. B. Shroff

Manuscript under review · 2026

Democratizing LLM Training with the Edge: Challenges, Systems, and the Road Ahead

Z. Luo, J. Cai*, C. Le Denmat*, S. Nair*, Fatemeh Nourzad*, et al.

IEEE/ACM Transactions on Networking · Invited paper · Publication forthcoming

Related workshop contributions

*Co-primary authors with equal contribution in alphabetical order.

Education, Skills & Service

Education

PhD, Electrical & Computer Engineering

The Ohio State University

Advisor: Ness B. Shroff

2022–present · Expected May 2028

MS, Electrical & Computer Engineering

University of Tehran

2019–2022

BS, Electrical & Computer Engineering

University of Tehran

2014–2018

Technical experience

Programming & ML
Python, C/C++, PyTorch, Hugging Face, Weights & Biases
LLM post-training
PPO, GRPO, DPO, PEFT/LoRA, tool-use RL
Distributed training
FSDP, DeepSpeed, SLURM, CUDA, gRPC
Infrastructure
Linux, Git, Docker, AWS, Google Cloud

Service & selected recognition

Peer review
NeurIPS, NeurIPS Workshops, ICLR, AISTATS, NAACL, IEEE/ACM ToN, ICDCS2022–present
Organizing
Co-organizer, Experts in AI and Networking, AI-EDGE Institute2023–present
Teaching
Teaching assistant, Machine Learning, Reinforcement Learning, and Stochastic Processes, University of Tehran2017–2022
Community
Panelist, AI-EDGE Summer REU Program2025
Presentations
  • Presenter, Artificial Intelligence Modeling, Analysis, and Control of Complex Systems (AIMACCS) Workshop2026
  • Oral presentation, ACM SIGCOMM NAIC Workshop; only extended-abstract submission selected for an oral presentation2026
  • Selected presenter, Kraus Memorial ECE Graduate Poster Competition2025
Back to top