Hi, I’m a final-year PhD student at Texas A&M University, advised by Gaurav Pandey and Reza Langari. My research focuses on efficient and controllable video generation and world models. My broader interest is in physical AI, bridging the gap between video generation and models that understand and interact with the physical world.
Previously, I worked with Shu Kong and Zhengzhong Tu and interned at Rivian, where I worked on reinforcement learning and model robustness. I earned my bachelor’s degree in Mechanical Engineering from Politecnico di Torino.
I enjoy understanding how models work, questioning their limitations, and experimenting with new ideas.
I am seeking industry research positions in video generation, world models, and physical AI.
Email: taghavi.pardis@gmail.com
arXiv, 2026 · Under review
Approximates dense attention with sparse computation and residual correction, accelerating video generators and world models without retraining. Achieves 1.48–2.61× end-to-end speedups while preserving generation quality.
arXiv, 2026 · Under review
Enables users to revise actions during world-model generation without restarting sampling. A lightweight corrector updates intermediate representations to reflect the new actions, supporting both whole-chunk changes and edits that preserve an existing prefix.
arXiv, 2026 · Under review
Introduces physical frame rate, a measure of whether video motion unfolds at a physically plausible speed. A learned visual chronometer estimates this rate, enabling playback corrections that improve the perceived naturalness of generated motion.
IROS, 2024
Jointly estimates depth and semantic segmentation from a single image using a shared transformer architecture. Adversarial training refines both predictions and improves performance on Cityscapes and NYU Depth V2.
arXiv, 2026 · Under review
Separates high level driving reasoning from motion planning. A large vision language model guides a lightweight driving model through an interpretable intermediate representation. This design reduces adaptation cost and improves planning performance on nuScenes.
CVPR Workshops, 2026
Distills foundation models into a compact instance segmentation model using limited labeled data. Teacher adaptation and contrastive learning reduce pseudo label bias. The student is approximately 11× smaller and outperforms its adapted teachers.
Graduate Researcher · College Station, TX
AI Engineer Intern · Palo Alto, CA
AVA studies autonomous driving in rural environments. I led perception research and development for urban and rural driving, focusing on visual scene understanding and multi-sensor fusion using cameras, LiDAR, and IMU.
Website
Developed a monocular perception pipeline using SwinMTL for joint depth estimation, semantic segmentation, and downstream 3D object detection. Validated in simulation and through remote testing on a real autonomous vehicle at Mcity. Our team secured an NSF subaward and ranked among the top four teams awarded testing access.
CodePhD in Mechanical Engineering · GPA: 4.0/4.0
BS in Mechanical Engineering · 110/110, summa cum laude
TOPolito Scholarship recipient.
