Our paper, Idea First, Code Later: Disentangling Problem Solving from Code Generation in Evaluating LLMs for Competitive Programming, has been accepted to Findings of ACL 2026!

I’ll be attending ACL in San Diego to present the paper and meet with collaborators. If you’re interested in meeting, do send me a message!

Previously, sense representations demanded specialized architectures and from-scratch training. We solve this by showing that sense representations are inducible properties of any pretrained model.

My new paper, Sense Representations are Inducible Interfaces is now out on arXiv as a preprint.

Current multilingual models cram 100+ languages into one parameter space and call the result multilingual. We argue for a linguistically grounded alternative: model multilinguality as sense adaptation, where new languages anchor onto representations of meanings the model already knows.

My new paper, Multilinguality as Sense Adaptation, is now out as a preprint on arXiv.

Special thanks to Mila - Quebec AI Institute and McGill University for hosting me while working on this project!