Puneet K. Dokania

Mixture of Experts Made Intrinsically Interpretable

RanDumb: Random Representations Outperform Online Continually Learned Representations

What Makes and Breaks Safety Fine-tuning? A Mechanistic Study

Fine-tuning can cripple foundation models; preserving features may be the solution

Online Continual Learning Without the Storage Constraint

Segment, Select, Correct: A Framework for Weakly-Supervised Referring Segmentation