Nikash Bhardwaj

Compositional Image Generation via Iterative Refinement

Shantanu Jaiswal*, Mihir Prabhudesai, Nikash Bhardwaj, Zheyang Qin, Amir Zadeh, Chuan Li, Deepak Pathak, Katerina Fragkiadaki
In submission
arxiv / code /

Designed a training-free, iterative self-correction pipeline for compositional text-to-image generation using a vision-language critic, improving all-correct rates by 16% on ConceptMix vs. compute-matched baselines.

Physics Neural Solver

Mihir Prabhudesai*, Aryan Satpathy*, Yangmin Li*, Zheyang Qin, Nikash Bhardwaj, Deepak Pathak, Katerina Fragkiadaki
In preparation
arxiv /

Post-training a large language model to solve olympiad-level physics problems using reinforcement learning and a physics simulator to construct synthetic question-answer pairs.

Recovering Walls and Cliffs from Monocular Images for Educational Robots

Nikash Bhardwaj, David Touretzky
CMU Meeting of the Minds, December, 2024
poster /

Performed computer vision research enabling low-cost educational robots to navigate using only onboard cameras and neural-network-based depth estimation, implementing edge detection, surface recovery, and world-map construction in Python so robots can recognize their surroundings and safely traverse unknown terrain.

Optimal Investment Management

Nikash Bhardwaj, Kathy Yan, William Hrusa
CMU Meeting of the Minds, May, 2024
poster /

Developed optimization models to solve agency and optimal investment problems for hedge funds, using Python with an Excel-based interface to choose fee structures that maximize the expected utility of both managers and investors under multi-period commission and utility functions.

	Compositional Image Generation via Iterative Refinement Shantanu Jaiswal, Mihir Prabhudesai, Nikash Bhardwaj, Zheyang Qin, Amir Zadeh, Chuan Li, Deepak Pathak, Katerina Fragkiadaki In submission* arxiv / code / Designed a training-free, iterative self-correction pipeline for compositional text-to-image generation using a vision-language critic, improving all-correct rates by 16% on ConceptMix vs. compute-matched baselines.
	Physics Neural Solver Mihir Prabhudesai, Aryan Satpathy, Yangmin Li, Zheyang Qin, Nikash Bhardwaj, Deepak Pathak, Katerina Fragkiadaki In preparation* arxiv / Post-training a large language model to solve olympiad-level physics problems using reinforcement learning and a physics simulator to construct synthetic question-answer pairs.
	Recovering Walls and Cliffs from Monocular Images for Educational Robots Nikash Bhardwaj, David Touretzky CMU Meeting of the Minds, December, 2024 poster / Performed computer vision research enabling low-cost educational robots to navigate using only onboard cameras and neural-network-based depth estimation, implementing edge detection, surface recovery, and world-map construction in Python so robots can recognize their surroundings and safely traverse unknown terrain.
	Optimal Investment Management Nikash Bhardwaj, Kathy Yan, William Hrusa CMU Meeting of the Minds, May, 2024 poster / Developed optimization models to solve agency and optimal investment problems for hedge funds, using Python with an Excel-based interface to choose fee structures that maximize the expected utility of both managers and investors under multi-period commission and utility functions.

Nikash Bhardwaj

Research

Compositional Image Generation via Iterative Refinement

Physics Neural Solver

Recovering Walls and Cliffs from Monocular Images for Educational Robots

Optimal Investment Management