Sahil Shah

I'm a first-year PhD student at University of Illinois Urbana-Champaign advised by Prof. Heng Ji in Blender Lab. Before that, I completed my BS and MS degrees at The University of Texas at Austin, advised by Prof. Sandeep Chinchali. I've also interned at ByteDance Seed, NVIDIA, and Tesla, working broadly on generative models and computer architecture.

Email / CV / Scholar / Github / LinkedIn

Research

I'm interested in building systems that learn structured, reusable representations of actions and scenes, with a focus on post-training vision-language models for long-video understanding and temporal reasoning. I'm also interested in world models for embodied reasoning and self-improving agents.

	CrossView: Can Vision-Language Models Reason Across Cameras? Sahil Shah, S. P. Sharan, Harsh Goel, Manvik Pasula, Adithya Hebbalae, Minkyu Choi, Sandeep Chinchali ECCV*, 2026 project page / arXiv / code A multi-camera video QA benchmark spanning driving, surveillance, ego/exo, and robotics, where today's best models still fall short of reasoning across simultaneous views.
	Incentivizing Vision Language Models to Search for Long Video Question Answering Harsh Goel, S. P. Sharan, Sahil Shah, Minkyu Choi, Joungbin An, Kristen Grauman, Sandeep Chinchali ECCV, 2026 (Spotlight Presentation) project page / arXiv / code Turning long-video QA into a multi-turn search, with temporal-logic-derived verifiable rewards for RL post-training, lifts Pass@1 by 8% and Pass@4 by 15%.
	ObjectAlign: Neuro-symbolic Object Consistency Verification and Correction Mustafa Munir, Harsh Goel, Xiwen Wei, Minkyu Choi, Sahil Shah, Kartikeya Bhardwaj, Paul Whatmough, Sandeep Chinchali, Radu Marculescu CVPRW, 2026 arXiv Verifying object consistency across frames catches objects that drift, morph, or vanish during video editing, then corrects them in place.
	RT-NeuS: Towards Real-Time Neuro-Symbolic Video Understanding via Adaptive Temporal Verification Shawn Liang, Sahil Shah, Chengwei Zhou, S. P. Sharan, Harsh Goel, Arnab Sanyal, Sandeep Chinchali, Gourav Datta NeuS, 2026 arXiv Adaptive sampling and keyframe selection bring neuro-symbolic temporal verification into the real-time regime without giving up multi-step reasoning.
	NeuS-QA: Grounding Long-Form Video Understanding in Temporal Logic and Neuro-Symbolic Reasoning Sahil Shah, S. P. Sharan, Harsh Goel, Minkyu Choi, Mustafa Munir, Manvik Pasula, Radu Marculescu, Sandeep Chinchali AAAI, 2026 project page / paper / code Training-free pipeline that identifies logical event sequences in video, boosting VQA accuracy by over 10% on causal and multi-step reasoning tasks.
	COFFEE: a High-Performance Approach to Convex Optimization for Thermodynamic Equilibrium Computations Fu-Yao Yu, Sahil Shah*, Yash Mittal, Paul Bessler, Aamir Mohsin, Jeffrey Geng, Arnav Vats, David Soloveichik SIEDS, 2025 (Best Paper)* project page / paper / code A trust-region convex solver for molecular equilibrium runs 2× faster and 10⁷× more accurately than prior tools, scaling to large biochemistry datasets.
	A Challenge to Build Neuro-Symbolic Video Agents Sahil Shah, Harsh Goel, Sai Shankar Narasimhan, Minkyu Choi, S. P. Sharan, Oguzhan Akcin, Sandeep Chinchali NeuS, 2025 arXiv / code Combining neuro-symbolic reasoning with video perception enables agents that can interpret, predict, and act on temporal events, not just recognize them.
	We'll Fix it in Post: Improving Text-to-Video Generation with Neuro-Symbolic Feedback Minkyu Choi, S. P. Sharan, Harsh Goel, Sahil Shah, Sandeep Chinchali arXiv preprint, 2025 arXiv Neuro-symbolic feedback enables zero-shot refinement of generated videos, boosting temporal and semantic alignment by nearly 40% without retraining.
	Neuro-Symbolic Evaluation of Text-to-Video Models using Formal Verification S. P. Sharan, Minkyu Choi, Sahil Shah, Harsh Goel, Mohammad Omama, Sandeep Chinchali CVPR, 2025 project page / paper / code Formally verifying a video against a temporal logic specification yields a text-to-video metric that aligns 5× better with human judgment than existing scores.
	Real-Time Privacy Preservation for Robot Visual Perception Minkyu Choi, Yunhao Yang, Neel P Bhatt, Kushagra Gupta, Sahil Shah, Aditya Rai, David Fridovich-Keil, Ufuk Topcu, Sandeep Chinchali TMLR*, 2025 arXiv Blurring objects based on logical specifications enables real-time video privacy with 95%+ compliance, provable guarantees, and seamless robot deployment.
	Towards Neuro-Symbolic Video Understanding Minkyu Choi, Harsh Goel, Mohammad Omama, Yunhao Yang, Sahil Shah, Sandeep Chinchali ECCV, 2024 (Oral Presentation) project page / paper / code Decoupling perception and temporal reasoning with TL-based state machines boosts long-range event identification by up to 15% on self-driving datasets.

Designed and based on Jon Barron's Website.