Privacy Protection and Reliable Evaluation in Multimodal AI

Multimodal large language models perform well on many visual and textual reasoning tasks, but assistive applications expose concerns about privacy and reliable real-world evaluation. This talk examines document understanding and spatial reasoning in three-dimensional environments. An agentic privacy-protection method masks personal information selectively while preserving visual context needed for question answering. SceneCritic, a symbolic evaluator, applies explicit spatial heuristics to test the semantic and geometric consistency of generated indoor layouts. Together, these projects investigate how multimodal models can be useful despite their limitations as reasoners and evaluators.