Di Wang

Current vision-language models can *see* point cloud defects, but can't reliably *diagnose* them, highlighting a critical gap in grounded quality understanding.

LLMs can now perform traceable, multi-step ecological reasoning over complex forest environments by operating on ecological hypergraphs and invoking deterministic tools, achieving higher accuracy and faithfulness than single-step approaches.

Despite showing promise in reading raw height data, today's MLLMs often fail to translate geometric perception into reliable semantic reasoning about natural scenes, even performing worse than RGB-only models when both modalities are needed.

Today's best language models can barely make sense of your messy group chats and fragmented digital life, achieving only 19% accuracy on a new benchmark of real-world reasoning.

ViTs can achieve robust generalization through adversarial training even when overfitting, mirroring a phenomenon previously observed only in CNNs.

Generate better peptide therapeutics faster: a new deep learning framework predicts peptide-protein interactions with high accuracy and generates novel peptides with enhanced binding affinity.

Training on a subset of data, a common technique for scaling ML, surprisingly introduces new privacy vulnerabilities by leaking information about both the training set and the selection process itself.

CAT's ability to defend against jailbreaks hinges on the singular values of the LLM's embedding matrix, offering a new handle for improving robustness.

LLMs, impressive as they are, can't juggle multiple users' conflicting needs without dropping balls on privacy, prioritization, and efficiency.