Soroush Mahdi

I’m Soroush Mahdi, an AI researcher working on 3D computer vision, multimodal learning, and efficient deep learning systems. My recent research focuses on how visual models can build scalable and robust representations of the world from images, videos, and multimodal data.

I earned my M.Sc. in Artificial Intelligence from Amirkabir University of Technology (Tehran Polytechnic), graduating with a GPA of 4.0/4.0. My thesis focused on improving the robustness of deep neural networks against adversarial attacks under the supervision of Prof. Maryam Amir Mazlaghani.

Currently, I am a Research Assistant at the Autonomous & Intelligent Systems Lab (AISL), where I work on 3D vision and large-scale visual perception systems. My recent work includes Evict3R, a training-free token eviction framework for efficient streaming 3D reconstruction transformers, and MODE-TTA, a method for robust test-time adaptation in 3D vision-language models under distribution shifts.

Research Interests

• 3D Computer Vision
• Vision-Language Models & Multimodal Learning
• Embodied AI & World Models
• Medical Image Analysis
• Robust and Trustworthy AI

More broadly, I am interested in building visual systems that can perceive, reason, and adapt in complex real-world environments. I will be pursuing a Ph.D. in AI and Computer Vision to continue working on these problems.

If you are interested in research collaboration or discussion, feel free to reach out.