University of California San Diego
I know you are busy. Ask what you want to know about me :)
→ See how I made this chat-bot from scratch in my github.
In 2017, I earned a silver medal in Iran’s National Physics Olympiad. I went on to complete my B.S. at the University of Tehran in computer engineering and am now pursuing my PhD at UCSD, doing research on stuff that I love. Outside of academics, I enjoy chess, tennis, and playing guitar.
I work under the guidance of Professor Farinaz Koushanfar, focusing on learning-based optimization. I want to get AI to solve complex tasks and design systems that streamline the training process, making it more efficient and scalable. I’m driven by the challenge of optimizing model performance while reducing computational costs.
<aside>
MoEs are Stronger than You Think: Hyper-Parallel Inference Scaling with RoE: Zibakhsh, Soheil, Mohammad Samragh, Kumari Nishu, Lauren Hannah, Arnav Kundu, and Minsik Cho. "MoEs Are Stronger than You Think: Hyper-Parallel Inference Scaling with RoE." arXiv preprint arXiv:2509.17238 (2025).
MoE-PHDS: One MoE checkpoint for flexible runtime sparsity Hannah, Lauren, Soheil Zibakhsh, Kumari Nishu, Arnav Kundu, Mohammad Samragh Razlighi, Mehrdad Farajtabar, and Minsik Cho. "MoE-PHDS: One MoE checkpoint for flexible runtime sparsity." arXiv preprint arXiv:2509.23012 (2025).
ForTIFAI: Fending Off Recursive Training Induced Failure for AI Model Collapse Shabgahi, Soheil Zibakhsh, Pedram Aghazadeh, Azalia Mirhoseini, and Farinaz Koushanfar. "ForTIFAI: Fending Off Recursive Training Induced Failure for AI Models." arXiv preprint arXiv:2509.08972 (2025).
MergeGuard: Efficient Thwarting of Trojan Attacks in Machine Learning Models Shabgahi, Soheil Zibakhsh, Yaman Jandali, and Farinaz Koushanfar. "MergeGuard: Efficient Thwarting of Trojan Attacks in Machine Learning Models." arXiv preprint arXiv:2505.04015 (2025).
LayerCollapse: Adaptive compression of neural networks: Shabgahi, Soheil Zibakhsh, Mohammad Soheil Shariff, and Farinaz Koushanfar. "LayerCollapse: Adaptive compression of neural networks." arXiv preprint arXiv:2311.17943 (2023).
LiveTune: Dynamic Parameter Tuning for Feedback-Driven Optimization: Shabgahi, Soheil Zibakhsh, et al. "LiveTune: Dynamic Parameter Tuning for Training Deep Neural Networks." arXiv preprint arXiv:2311.17279 (2023).
Modeling Effective Lifespan of Payment Channels: Shabgahi, Soheil Zibakhsh, et al. "Modeling Effective Lifespan of Payment Channels." arXiv preprint arXiv:2301.01240 (2022).
****Throughput limitation of the off-chain payment networks: Dehshali, Shayan Hamidi, et al. "Throughput limitation of the off-chain payment networks." Cryptology ePrint Archive (2022).
Internship at Apple Spring and Summer 2025