Mukilan Karuppasamy

I'm a Research Scientist at FastcodeAI, where I built Hydra, a speech-to-speech model for SmallestAI.

My research spans multimodal LLMs, interpretability, and robustness. Previously, I was a predoctoral researcher at IISc Bangalore with Prof. Anirban Chakraborty, working on visual instruction tuning for MLLMs. Before that, I built the first interpretable video action classifier with Prof. C.V. Jawahar (IIIT Hyderabad), large-scale noise robustness analysis on vision-language models with Prof. Yogesh Rawat (CRCV, UCF), and explored distributed message passing as a MITACS Globalink Fellow with Prof. Amine Mezghani and Prof. Faouzi Bellili (University of Manitoba, 2023).

Email  /  Resume  /  Scholar  /  Github  /  Linkedin

Research

I'm interested in mechanistic interpretability, representation learning, and robustness analysis. My current interest lies in understanding the theoretical basis of interpretability in overparameterized deep neural networks. Some papers are highlighted.

Comprehensive analysis of VLM/LLM across tasks on number representations when numbers are split across tokens.

We critically examine the vulnerability of SOTA Zero-Shot Object Detectors with 50 different backbones to resolution degradation (`pixelation'). Three key takeaways are 1) The choice of backbone alone can substantially enhance robustness, while prompting (language) plays a minimal role. 2) Robustness is also a function of the dataset/domain of images at inference. 3) Cross-exchange of vision features within backbone helps in improving robustness.

Proposed a large scale sketch based pretraining and visual instruction tuning dataset with 30M instances. Developed a state-of-the-art sketch-based multimodal model that can perform detection, Counting, VQA, and SBIR with any crude, hand-drawn sketches as the query instead of natural language.

Proposed a concept based inherently intrepretable model for Video models using concept bottleneck model and token merging. This was applied in Driver Intention Prediction task.

Derived a collaborative signal estimation from multiple agents with different measurement channels through distributed message passing algorithm.

Services

Reviewer - AAAI 2026, AAAI 2027, ICLR 2027