Hexuan (Matt) Wang

I am interested in Natural Language Processing and Multimodal Intelligence, with a focus on making models reliable. My research asks: when a model tells us something, can we trust it and verify it? A model should be grounded, anchoring each claim in evidence, whether a cited passage, an image region, or a piece of code; verifiable, so that a person can check its outputs, and check them efficiently; and faithful to its own knowledge, abstaining rather than hallucinating when evidence is lacking. I study these properties in attributed generation, over scientific tables, under temporal shift, and across modalities, via rigorous evaluation, post-training, and neuro-symbolic methods. Ultimately, I aim to build trustworthy models for scientific discovery and high-stakes decisions.