PhD student, University of Wisconsin-Madison
1 paper at NeurIPS 2025
We propose Vittle, a new visual instruction tuning framework that improves robustness of MLLMs to data distribution shifts by pursuing the minimal sufficient representation.
