Michal Shlapentokh-Rothman

I am a 5th year computer science PhD student at the University of Illinois at Urbana-Champaign. I am co-advised by Professors Derek Hoiem and Yuxiong Wang. Previously, I received my bachelor's and Masters of Engineering degrees in computer science from MIT in 2019 and 2020. At MIT, I was a member of the ALFA Group .

I am looking for research internships for summer 2025.

Email /  Google Scholar  /  LinkedIn  /  CV

Dec. 2024: Can We Generate Visual Programs Without Prompting LLMs? released. The answer is yes and only requires 500 question/answer pairs!
June 2024: Region representations presented at CVPR 2024.
May 2023: WebWISE released. Investigates using GPT-3 for web tasks
July 2021: Our Learning Curves paper was presented at ICML.
May 2021: I will be interning at Amazon this summer

My overall research goal is to create adaptable and efficient visual agent systems, where foundation models can be composed and trained for a wide variety of downstream tasks with limited user input. During my undergrad, I did research on using evolutionary algorithms for cyber security.

We develop a synthetic data augmentation approach and alternative program generation method based on decoupling programs into higher-level skills called templates and the corresponding arguments. Our results show that with data augmentation, prompt-free smaller LLMs (approx. 1B parameters) are competitive with GPT-4o-mini on VQA tasks with the added benefit of much faster inference.

Region representations used to be popular in the pre-deep learning era. What happens when we create region representations with recently released founation models? We show that our simple method achieves impressive performance on existing tasks such as semantic segmentation as well as new one.

Combines LLM techinques for differnet reasoning and planning benchmarks.

How can we use GPT-3 for web-based tasks? We investigate the performance of GPT-3 on the Mini-WoB dataset uisng Document Object Model (DOM) elements as part of the input.

In this work, we use learning curves to investigate the effects of various design choices on neural network performance. We propose a method for estimating learning curves, abstract their parameters into error and data-reliance, and evaluate the effectiveness of different parameterizations.

In this work, we use co-evolutionary algorithms to explore the dynamics between cyber attack patterns and potential mitigations.

In this work, we show how we can use a competitive co-evolutionary framework to evaluate different software defined perimeters (SDP).

Artificial Intelligence (CS 440): Fall 2020
Computational Photography (CS 445): Spring 2021, Spring 2023

My Father
My Mother
My Brother
Gauss, the dog (11/21/06-5/31/19)
Markov, the dog
Hilbert, the dog

Website based on template from Jon Barron