Home

Biography

Shivvrat Arya is an Assistant Professor of Computer Science at the Ying Wu College of Computing, New Jersey Institute of Technology (NJIT), Director of the ARIA Lab, and a Core Faculty member of the Center for AI Research. His research develops structured, interpretable, and efficient AI methods that integrate learning, reasoning, and optimization, with a focus on probabilistic and neurosymbolic reasoning, neural combinatorial optimization, structured and multimodal intelligence, and AI for scientific discovery.

He received his Ph.D. in Computer Science from The University of Texas at Dallas, where he was advised by Vibhav Gogate and Yu Xiang. His doctoral research focused on learning-based methods for probabilistic inference and structured reasoning.

Research Focus

• Neuro-Symbolic and Probabilistic Reasoning: Integrating deep learning with probabilistic graphical models, circuits, and classical solvers to enable scalable, guaranteed, and real-time inference under uncertainty and constraints.
• Neural Combinatorial Optimization: Developing deep reinforcement learning and graph representation learning methods to learn adaptive, structure-aware decision policies for complex combinatorial and network optimization problems.
• Structured and Multimodal Intelligence: Bridging high-dimensional perception with procedural workflows, temporal reasoning, and human-in-the-loop guidance for procedural video understanding and trustworthy multimodal systems.
• AI for Scientific Discovery: Incorporating domain priors, biological networks, and mechanistic constraints into deep generative models for structured scientific discovery, with an emphasis on single-cell genomics and computational biology.

Research Highlights

• Research Recognition: Publications recognized with best paper awards, spotlights, and oral presentations at top AI/ML venues
• Neural-Augmented Probabilistic Inference: Developed NeuPI, an open-source neural inference engine accelerating probabilistic reasoning from minutes to microseconds, alongside Neural Dual Bounds (NeurIPS 2026 Spotlight, top 1%) and Learning to Condition (NeurIPS 2025).
• Neural Combinatorial Optimization: Developed RELINK (CIKM 2025), a deep reinforcement learning framework for sequential edge activation and influence maximization in privacy-constrained closed networks.
• Structured Multimodal Intelligence: Released CaptainCook4D (NeurIPS 2024 D&B Track), a 94.5-hour egocentric 4D dataset for procedural activity understanding, error recognition, and assistive AR guidance.
• AI-Driven Augmented Reality Guidance: Developed real-time AR guidance systems for complex physical tasks, combining predictive modeling with proactive assistance for multi-step procedures.

For earlier externally funded research contributions, including work supported through DARPA, NSF, and AFOSR projects, see Funded Projects.