Vin Howe

I’m a MATS 10.0/10.1 scholar supervised by Oliver Sourbut, where I'm working on a benchmark for multi-agent epistemic propensities. I'm also a Master's student in computer science at Brigham Young University; under the supervision of David Wingate, I study learning mechanics in LLMs. Previously, I did my Bachelor's in BYU's Applied and Computational Mathematics program.

Research Interests

I'm interested in deeply understanding AI to prevent catastrophic risks. In particular, I've researched learning mechanics to help predict how AI (mis)generalizes, and, as part of MATS, I've benchmarked multi-agent epistemics to help determine if agents are responsible enough to be trusted with our knowledge commons. Previously, I worked on pro-social applications of AI, like creating a chatroom that used AI suggestions to help improve online political conversations.

Blog

You probably shouldn’t train a language model in your browser—here’s how

Nov 21, 2025

Publications

2026

Preview of compartmentalization preprint Language models struggle with compartmentalization
Thomas V Howe, David Wingate
Preprint

2024

Diagram of gradient sparse autoencoder (gSAE) Features that Make a Difference: Leveraging Gradients for Improved Dictionary Learning
Jeffrey Olmo, Jared Wilson, Max Forsey, Bryce Hepner, Thomas V Howe, David Wingate
Findings of the Association for Computational Linguistics

2023

Preview of chatroom Leveraging AI for democratic discourse: Chat interventions can improve online political conversations at scale
Lisa P Argyle, Christopher A Bail, Ethan C Busby, Joshua R Gubler, Thomas V Howe, Christopher Rytting, Taylor Sorensen, David Wingate
Proceedings of the National Academy of Sciences

Projects

2025

I created Sequence Toy, a web playground for training small language models with WebGPU.

2021

I wrote the software the drives “The Wall,” the floor-to-ceiling interactive display in the lobby of BYU’s computer science building.