Prabhjot Singh

VLMs are failing to deliver equitable performance across different scripts, with accuracy gaps of up to 16% that could hinder access for billions of users.

Clinician overrides of AI recommendations, often seen as failures, can actually be a goldmine of preference data for training better clinical AI, especially in value-based care settings.

Image generators still spewing unsafe content? ReVision offers a training-free "last line of defense" that surgically removes policy violations *after* generation, without wrecking image quality.

Circuit cutting introduces substantial end-to-end overheads in quantum neural network training, with reconstruction dominating per-query time, but surprisingly, test accuracy and robustness can be preserved or even improved.