Talk notes, tutorials, and research experiments, written to make the central question, approach, and takeaways easy to follow. For full papers and publication details, visit the publications page.
Pilot experiments, Research notes, Riverside, CA
A small experiment on whether fine-tuning a vision-language model for one visual task helps it handle other tasks on the same images.
Invited talk, Mila Multiomics Reading Group, Montreal, Canada
A walkthrough of dual NUC/BPE tokenization for RNA, why truncation is a bad default, and how a 117M-parameter encoder can cover both short structural tasks and arbitrarily long sequences.
Tutorial, ACL 2024, Bangkok, Thailand
A tutorial on how safety-aligned language and multimodal models still fail under adversarial pressure, organized around text-only attacks, multimodal attacks, causes, and defenses.