Utkarsh Mishra

Email  |  Google Scholar  |  GitHub  |  LinkedIn  |  X

Hi, I am Utkarsh Mishra, an MS student in Computer Engineering at Texas A&M University. My research interests lie in Large Language Models (LLMs), Multimodal LLMs, agentic AI systems, reinforcement learning from human feedback (RLHF), and vision-language understanding.

I graduated with a B.Tech in Electrical Engineering and a minor in Computer Science from IIT Gandhinagar. Previously, I worked as a System Analyst at HDFC Bank.

I'm passionate about building intelligent systems that can reason, plan, and interact with the world through multimodal understanding. Let's connect to discuss innovative ideas in these spaces!

Industrial and Research Experience

News & Achievements

• Jan 2026: Paper on Constructive Distortion accepted at International Conference on Learning Representations (ICLR) 2026.
• Jan 2026: Paper on City Navigation accepted at European Chapter of the Association for Computational Linguistics (EACL) 2026 with Oral Presentation.
• Oct 2025: Started as Research Engineer at Texas A&M University (Dept. of AGLS).
• Aug 2025: Started MS in Computer Engineering at Texas A&M University.
• May 2025: Paper on 3D Point Cloud Denoising using Diffusion Transformers accepted at International Conference on Image Processing (ICIP) 2025.
• Mar 2025: Received admit from Texas A&M University with ECEN Merit Scholarship.
• Jun 2024: Graduated from IIT Gandhinagar.

Publications

City Navigation in the Wild: Exploring Emergent Navigation from Web-Scale Knowledge in MLLMs

Constructive Distortion: Improving MLLMs with Attention-Guided Image Warping

A novel approach combining transformer architectures with multi-resolution hash encoding in diffusion models for effective 3D point cloud denoising.

Projects

Developed a visual grounding module for autonomous web agents using Direct Preference Optimization (DPO) to train a Qwen2-0.5B model for predicting optimal screenshot crop regions. Integrated into SeeAct framework with LiteLLM support for GPT-4o, Claude, Gemini, LLaVA.

Evaluated whether video generation models (Sora, Veo3, Grok, Wan2.1) can produce physically realistic data for robotics training. Designed physics benchmark suite testing gravity, buoyancy, and collision dynamics with SAM2-based kinematics analysis.

Extended 3D Gaussian Splatting with denoising techniques (DBSCAN, point-wise distance pairing on input point cloud) and novel regularization techniques to reduce visual artifacts when using low resolution wide angle images.

Used GAN inversion on StyleGAN to analyze effect of object rotation on latent representation and generate novel views through latent space manipulation.

Combined Graph Convolutional Networks with LSTM to model spatial relationships and temporal dynamics for pose classification using OpenPose and AlphaPose.

Generated high-quality synthetic images using StyleGAN-XL and Stable Diffusion for CIFAR-10 dataset . Analyzed effects on classifier accuracy with varying synthetic to real data ratios.

Developed wearable device using STM32 Nucleo microcontroller and flex sensors. Implemented USB CDC protocol with Scikit-learn MLP classifier for gesture detection.