Yusong Wang

Instead of stuffing raw context with dialogue history or brittle retrieval vectors, treating user personalization as particle-filtered hypothesis tracing over natural language resolves the tension between fleeting intent and long-term preferences.

Misclassifications in biomedical imaging can be drastically reduced with BioMedVR's confusion-aware approach, leading to superior model performance in critical healthcare applications.

Reversing input text can unlock hidden contextual information, leading to richer embeddings and substantial performance gains in training-free LLMs.

VLMs can achieve state-of-the-art adversarial robustness by iteratively refining visual and textual representations through a closed-loop prompting mechanism, even with frozen encoders.