Research

Calvin and Hobbes

My interests are in scalable training objectives and adaptive inference methods for large-scale generative models. My work bridges foundational machine learning with strong empirical performance to scale inference, enhance capabilities, and improve the efficiency of multimodal agents across language, vision, and time-series data.

Concepts, Compositions, and Counterfactuals: Machine Abstractions for Human-Like AI
Bhishma Dedhia
PhD Thesis, Princeton University
[thesis]
Bottom-up Domain-specific Superintelligence: A Knowledge Graph is What We Need
Bhishma Dedhia, Yuval Kansal, Niraj K Jha
Preprint
[website] [arXiv] [pdf] [code]
Generating, Fast and Slow: Scalable Parallel Video Generation with Video Interface Networks
Bhishma Dedhia, David Bourgin, Krishna Kumar Singh, Yuheng Li, Yan Kang, Niraj K Jha, Yuchen Liu
International Conference on Computer Vision (ICCV), 2025
[website] [arXiv] [pdf]
Neural Slot Interpreters: Grounding Object Semantics in Emergent Slot Representations
Bhishma Dedhia, Niraj K Jha
Transactions of Machine Learning Research (TMLR), 2025
[arXiv] [pdf] [code]
Zero-TPrune: Zero-Shot Token Pruning through Leveraging of the Attention Graph in Pre-Trained Transformers
Hongjie Wang, Bhishma Dedhia, Niraj K Jha
Conference on Computer Vision and Pattern Recognition (CVPR), 2024
[website] [arXiv] [pdf]
Im-Promptu: In-Context Composition from Image Prompts
Bhishma Dedhia, Michael Chang, Jake C. Snell, Thomas L. Griffiths, Niraj K. Jha
Neural Information Processing Systems (NeurIPS), 2023
[website] [arXiv] [pdf] [code]
SCouT: Synthetic Counterfactuals via Spatiotemporal Transformers for Actionable Healthcare
Bhishma Dedhia*, Roshini Balasubramanian*, Niraj K Jha
ACM Transactions on Computing for Healthcare 2023
[arXiv] [pdf] [code]
FlexiBERT: Are Current Transformer Architectures too Homogeneous and Rigid?
Shikhar Tuli, Bhishma Dedhia, Shreshth Tuli, Niraj K Jha
Journal of AI Research 2023
[arXiv] [pdf] [code]
Whittle Index based Age-of-Information Aware Scheduling for Markovian Channels
B Sombabu, Bhishma Dedhia, Sharayu Moharir
Computer Networks and Communications 2023
[wiserpub] [pdf]
Saliency-driven rate-distortion optimization for 360-degree image coding
Jui-Chiu Chiang, Cheng-Yu Yang, Bhishma Dedhia, Yi-Fan Char
Multimedia Tools and Applications 2021
[springer]
Lower Bounds for Policy Iteration on Multi-action MDPs
Kumar Ashutosh*, Sarthak Consul*, Bhishma Dedhia*, Parthasarathi Khirwadkar*, Sahil Shah*, Shivaram Kalyanakrishnan
Conference on Decision and Control 2020
[arXiv] [pdf] [code]
On Minimizing Channel-Aware Age of Information in a Multi-Sensor Setting
Bhishma Dedhia, Sharayu Moharir
IIT-Bombay Senior Thesis
[arXiv] [pdf]
You Snooze, You Lose: Minimizing Channel-Aware Age of Information
Bhishma Dedhia, Sharayu Moharir
International Symposium on Modeling and Optimization in Mobile, Ad Hoc and Wireless Networks (WiOpt) 2020
[arXiv] [pdf]
Saliency Prediction for Omnidirectional Images Considering Optimization on Sphere Domain
Bhishma Dedhia, Jui-Chiu Chiang, Yi-Fan Char
International Conference on Acoustics, Speech and Signal Processing (ICASSP) 2019
[arXiv] [pdf]
Analysis of Lower Bounds for Simple Policy Iteration
Sarthak Consul*, Bhishma Dedhia*, Kumar Ashutosh*, Parthasarathi Khirwadkar*
Technical Report 2019
[arXiv] [pdf] [code]

* Equal contribution

Also on Google Scholar