arXiv subjects

Utkarsh Maurya

Publications and source records attributed to Utkarsh Maurya.

6 recordsLinked to original sources

arXiv2026

OMIND: Framework for Knowledge Grounded Finetuning and Multi-Turn Dialogue Benchmark for Mental Health LLMs

Large Language Models (LLMs) have shown remarkable capabilities for complex tasks, yet adaptation in medical domain, specifically mental health, poses specific challenges. Mental health is a rising concern globally with LLMs having large potential to help address the same. We highlight three primary challenges for LLMs in mental health - lack of high quality interpretable and knowledge grounded training data; training paradigms restricted to core capabilities, and evaluation of multi turn dialogue settings. Addressing it, we present oMind framework which includes training and aligning LLM agents for diverse capabilities including conversations; high quality ~164k multi-task SFT dataset, as a result of our generation pipeline based on Structured Knowledge retrieval, LLM based pruning, and review actions. We also introduce oMind-Chat - a novel multi turn benchmark dataset with expert annotated turn level and conversation level rubrics. Our diverse experiments on both core capabilities and conversations shows oMind LLMs consistently outperform baselines. oMind-LLM also shows significantly better reasoning with up to 80% win rate.

cs.CL↗

arXiv2026

Optimal transport and control of an active particle near a plane wall

The control of active colloidal particles via optical traps is a cornerstone for research of matter at the micron and nanometer scale. A central challenge in this domain is the derivation of optimal transport protocols that minimize the mean work required to move a particle over a finite-time interval. Here, we present the Ritz method in which open-loop protocols are constructed from a global basis of Chebyshev polynomials. The protocols are optimized using either a genetic algorithm or a gradient-based method. We apply the method to study optimal transport of an active particle, which is modeled as a force-dipole (or a stresslet) near a no-slip wall. The methodology is validated in the limits of zero activity and infinite wall separation, where it successfully recovers the known analytical protocols and the theoretical minimum work. Crucially, we demonstrate that the presence of the activity breaks the time-reversal symmetry of the optimal protocol found. This symmetry breaking is shown to be a complex function of the transport direction and the particle's intrinsic activity. Because the presented approach requires only the capability to simulate stochastic trajectories, it offers a robust, principled framework for optimizing transport protocols in complex fluid environments that remain inaccessible to exact analytical treatment.

cond-mat.soft↗

arXiv2025

Post-Pruning Accuracy Recovery via Data-Free Knowledge Distillation

Model pruning is a widely adopted technique to reduce the computational complexity and memory footprint of Deep Neural Networks (DNNs). However, global unstructured pruning often leads to significant degradation in accuracy, typically necessitating fine-tuning on the original training dataset to recover performance. In privacy-sensitive domains such as healthcare or finance, access to the original training data is often restricted post-deployment due to regulations (e.g., GDPR, HIPAA). This paper proposes a Data-Free Knowledge Distillation framework to bridge the gap between model compression and data privacy. We utilize DeepInversion to synthesize privacy-preserving ``dream'' images from the pre-trained teacher model by inverting Batch Normalization (BN) statistics. These synthetic images serve as a transfer set to distill knowledge from the original teacher to the pruned student network. Experimental results on CIFAR-10 across various architectures (ResNet, MobileNet, VGG) demonstrate that our method significantly recovers accuracy lost during pruning without accessing a single real data point.

cs.LG↗

arXiv2023

Evaluating the anticipated outcomes of MRI seizure image from open-source tool- Prototype approach

Epileptic Seizure is an abnormal neuronal exertion in the brain, affecting nearly 70 million of the world's population (Ngugi et al., 2010). So many open-source neuroimaging tools are used for metabolism checkups and analysis purposes. The scope of open-source tools like MATLAB, Slicer 3D, Brain Suite21a, SPM, and MedCalc are explained in this paper. MATLAB was used by 60% of the researchers for their image processing and 10% of them use their proprietary software. More than 30% of the researchers use other open-source software tools with their processing techniques for the study of magnetic resonance seizure images

physics.med-ph↗

arXiv2023

India's Progress in Space Exploration and International Legal Challenges in Meeting Goals within International Space Boundaries: A Review

India's journey towards space has been nothing short of remarkable. It began with the establishment of the Indian Space Research Organization (ISRO) in 1969, with the aim of harnessing space technology for national development. Since then, India has made significant strides in space exploration, satellite technology, and space research. One of the major milestones in India's space journey was the launch of the first Indian satellite, Aryabhata, on April 19, 1975. This marked India's entry into the exclusive club of space-faring nations. The success of Aryabhata paved the way for further advancements in space technology and inspired a generation of scientists and engineers.

physics.pop-ph↗

arXiv2023

Detection and Classification of Glioblastoma Brain Tumor

Glioblastoma brain tumors are highly malignant and often require early detection and accurate segmentation for effective treatment. We are proposing two deep learning models in this paper, namely UNet and Deeplabv3, for the detection and segmentation of glioblastoma brain tumors using preprocessed brain MRI images. The performance evaluation is done for these models in terms of accuracy and computational efficiency. Our experimental results demonstrate that both UNet and Deeplabv3 models achieve accurate detection and segmentation of glioblastoma brain tumors. However, Deeplabv3 outperforms UNet in terms of accuracy, albeit at the cost of requiring more computational resources. Our proposed models offer a promising approach for the early detection and segmentation of glioblastoma brain tumors, which can aid in effective treatment strategies. Further research can focus on optimizing the computational efficiency of the Deeplabv3 model while maintaining its high accuracy for real-world clinical applications. Overall, our approach works and contributes to the field of medical image analysis and deep learning-based approaches for brain tumor detection and segmentation. Our suggested models can have a major influence on the prognosis and treatment of people with glioblastoma, a fatal form of brain cancer. It is necessary to conduct more research to examine the practical use of these models in real-life healthcare settings.

eess.IV↗