Extended frame-based concept guidance into vision-language models and used it to inspect safety weaknesses in Qwen2-VL and Llama 3.2 Vision.

Modeled frames so multi-token words can be measured, compared, and used for concept-guided generation across modern LLMs.

Evaluated humor detection in stand-up comedy by comparing LLM-selected punchlines with audience-laughter ground truth.

Framed CSAM triage with law-enforcement guidance as a safer proxy for classifying abuse imagery scenes.

Generalized first-order difference subspaces to capture velocity and acceleration-like dynamics on Grassmann manifolds.

Designed OSA-DAS, a perturbation-based visual explanation method that compares augmentation and occlusion subspaces in deep feature space.