Hong Yu

Current vision-language models can generate fluent medical explanations, but struggle to ground them in relevant visual evidence, especially in emotionally charged scenarios.

Ditch the slow external knowledge base: this new method bakes clinical knowledge directly into LLMs for lightning-fast, state-of-the-art healthcare predictions.

Achieve near-lossless 4-bit quantization for LLMs in under a minute, without full fine-tuning, by correcting for non-uniform activation distributions.

Medical VQA gets a boost: Mamba-based models enhanced with knowledge graphs now outperform existing methods by better associating lesion features with diagnostic criteria.

By prioritizing gradient direction in token optimization and using a two-stage loss, TAO-Attack achieves near-perfect jailbreak success rates against multiple LLMs, exposing critical vulnerabilities in current safety alignments.

Clinical question answering gets a boost: TARSE aligns language model reasoning with clinically valid logic by retrieving and adapting to relevant skills and prior reasoning experiences at test time.

Overcome underwater visibility challenges and improve fish detection accuracy by explicitly modeling and compensating for frequency-specific information loss due to wavelength-dependent absorption and turbidity.