Mark Ibrahim

For datasets with high-magnitude noise features, joint-embedding is more robust than reconstruction for self-supervised learning.

we release a cognitively-inspired benchmark for reasoning across scenes that reveals hallucination is an open challenge for multimodal models

Increased verbatim memorization doesn't necessarily lead to greater chat extractability, and model quality is a greater privacy threat than memorization

Our new benchmark AbstentionBench reveals reasoning models struggle to determine when not to answer.