Uncertainty Quantification of MLLM-as-a-Judge (ASU)

Studying uncertainty in multimodal LLM-based evaluators across pointwise, pairwise, and batch evaluation settings, with a focus on reliability, calibration, and consistency of judgment in automated evaluation pipelines.

Localized Concept Erasure in Generative Models (ASU)

Developing methods for fine-grained concept erasure in diffusion and flow-matching models, addressing limitations of prior approaches (e.g., single-concept training and failure under multiple instances), alongside designing evaluation metrics that distinguish true unlearning from concealment and support multi-granular assessment.

Uncertainty Quantification of MLLM-as-a-Judge (ASU)

Mentored Developing a framework for AI authorship transparency and assessment integrity using document-level watermarking and adversarial PDF perturbations to study and identify AI-assisted usage in educational settings. This work was featured and being used at ASU.