Chaim Baskin

Unfrozen video-language models implicitly know when they've seen enough evidence, a signal that can be extracted and used to boost answer accuracy by nearly 10% without retraining.

CXR-Retrieve reveals that effective chest X-ray retrieval hinges on understanding not just what findings are present, but how they are clinically asserted, leading to substantial gains in precision.

Sparse, task-aware radar configurations can outperform traditional full-array setups, revolutionizing radar design for autonomous vehicles.

Weight inconsistency can be bounded for efficiency gains, leading to training time improvements of up to 1.69x without sacrificing stability.

Get GPT-4o-level long-video QA performance with 10x fewer FLOPs by using a hierarchical, training-free frame selector that combines multimodal experts and fuzzy logic.