Vatsal Gupta
MS Machine Learning · CMU

I am learning ML to try and build AI models that hold up.

I'm a Master's student in Machine Learning at Carnegie Mellon University. One question runs through my work: does a model actually understand the structure in its data, or did it find a correlation that happened to hold where it was tested?

I've explored it by stress-testing language and vision-language models, and by building physical structure into generative models, with Prof. Min Xu (CMU) and Prof. Dan Roth & Prof. Vivek Gupta (UPenn). Before CMU, I spent two years as a quant researcher at AlphaGrep Securities, where models had to keep working with real capital on the line. My Master's is supported by the Narotam Sekhsaria and K.C. Mahindra fellowships.

Email CV Scholar GitHub LinkedIn
Vatsal smiling in front of a window at the Palace of Versailles

At a glance

  • NowMS in Machine LearningCarnegie Mellon University · '27
  • PreviouslyQuant ResearcherAlphaGrep Securities · 2024–26
  • UndergradB.Tech in CSEIIT Guwahati
  • Thinking aboutWhen does a model really reason?Robustness under shift · multimodal reasoning · physical priors
  • FellowshipsNarotam Sekhsaria · K.C. MahindraSupporting my MS at CMU
Robustness Multimodal reasoning NLP Computer vision Flow matching

News

  1. Started my MS in Machine Learning at CMU.

  2. MicroFM, my work with the Xu Lab at CMU, was accepted to CVPR 2026.

  3. Wrapped up two years as a Quant Researcher at AlphaGrep.

  4. Presented NTSEBench at NAACL 2025.

  5. Presented Multi-Set Inoculation at EMNLP 2024.

  6. Graduated third in CSE at IIT Guwahati and started as a quant at AlphaGrep, building alphas for NSE, CME and MCX markets!

Experience

  1. Quant Researcher · AlphaGrep Securities

    Research on systematic strategies across options and futures, where a signal has to keep holding as markets shift. Built an LLM-based news-sentiment signal and a config-driven framework for taking alpha ideas live.

  2. Research Intern · Xu Lab, CMU

    Can a generative model recover 3D microscopy volumes without inventing detail? Worked on physics-based priors for MicroFM, a flow-matching reconstruction model.

  3. Research Intern · Cognitive Computation Group, UPenn

    Do language and vision-language models reason, or pattern-match? Built benchmarks and robustness methods to find out.

  4. ML Research Intern · Imperial College London

    Unsupervised autoencoders for real-time anomaly detection on a carbon-capture pilot plant.

Selected research

All publications →