Mukul Ranjan

I am a first-year PhD student in ECE at Cornell University, where I work with Prof. Udit Gupta on hardware-software co-design. My research focuses on efficient machine learning systems, particularly developing alternate architectures and hardware-aware algorithms through sparsity and quantization. I also work on creating evaluation methodologies and benchmarks for these systems.

Previously, I completed my MS in Machine Learning at MBZUAI, where I worked with Prof. Zhiqiang Shen, Prof. Praneeth Vepakomma, and Prof. Nils Lukas. I also worked with Prof. Deming Chen at UIUC from Jul. 2025 to Oct. 2025. Before that, I was a Data Scientist at Meesho, where I built personalized ranking systems, and an AI Research Scientist at Qure.ai, developing automated stroke severity assessment systems deployed in hospitals worldwide.

I hold a B.Tech. in Electronics and Communication Engineering from IIT Guwahati.

Email  /  CV  /  Scholar  /  Github  /  LinkedIn  /  X  /  Medium  /  Blogs  /  Poetry

News

Publications

Representative papers are highlighted. * indicates equal contribution (random order).

Flash-dLLM is a training-free framework that pairs an I/O-aware fused KV-cache kernel with draft-and-verify parallel decoding, achieving 5.1x and 11.0x speedups over Elastic-Cache on GSM8K and HumanEval.

Elastic-Cache is a training-free framework that accelerates Diffusion LLMs up to 45.1x with higher accuracy by adaptively refreshing the KV cache.

SpookyBench reveals that patterns in temporal noise that humans recognize with 98% accuracy, state-of-the-art VLMs fails completely achieving 0%.

Mobile-MMLU is a benchmark with with 16,000+ questions across 80 mobile-related fields to evaluate LLM performance under real-world constraints.

GBLM-Pruner is a gradient-based pruning method that is extremeley faster than weight-update methods like SparseGPT.

TAB-VLM is a benchmark of 600 questions over 1,600 Indian cultural artifacts that exposes cultural anachronism in VLMs; even the best model reaches only 58.7% accuracy.

CASS is the first large-scale dataset and model suite for cross-architecture GPU code transpilation (CUDA to HIP, Nvidia SASS to AMD RDNA3), with 70k verified code pairs.

GLoRA is a unified PEFT framework achieving state-of-the-art accuracy with zero inference overhead through structural re-parameterization.

KITAB-Bench has 8,809 samples across 9 domains. It reveals that vision-language models outperform traditional OCR by 60%.

Deep-ASPECTS is an automated ASPECT scoring system achieving radiologist-level performance, now deployed in hospitals worldwide.

Last updated: September 2026
© 2026 Mukul Ranjan. Design inspired by Jon Barron.