Research
google scholar ↗Two projects so far, and they look unrelated until you line them up. One asks whether a model can be argued out of a wrong answer. The other asks what a model loses when you teach it something new. Both are really about trusting a system you cannot inspect directly.
undergraduate researcher · jan 2026 — present
Continual customization for diffusion-based text-to-video generation. The setup is simple to state and annoying to achieve: a model learns a new subject, then another, then another, and it should still render the first one correctly at the end.
I am prototyping architecture-level extensions to DreamVideo and ModelScopeT2V-style pipelines. Subject conditioning, adapter-based customization, and temporal modules over denoising U-Net video backbones.
The measurements are all of damage rather than improvement. Catastrophic forgetting, identity drift, whether the model still listens to the prompt, and whether motion stays coherent across sequential customization tasks. Nothing published yet.
nsf reu, xinda wang lab · may — aug 2025
Ask one language model whether a C function is vulnerable and it returns a verdict plus a confident justification. It writes the justification either way. MAVUL starts from the position that a finding should have to survive an argument before it counts: a ReAct agent analyzes, a Reflexion agent attacks that analysis, and an evaluation agent judges whether either was right and records where the reasoning failed.
Evaluated on 879 real-world C/C++ vulnerable code pairs from PrimeVul against GPTLens, VulTrial and a single-agent GPT-4o baseline.