Nan Zhang

Achieving high-quality document restoration without manual prompts, DocPure outperforms traditional methods by leveraging degradation-aware structural insights.

Structured scene graphs can dramatically enhance MLLMs' reasoning capabilities, leading to substantial performance gains on complex visual tasks.

LLMs can learn to abstain from answering questions they're unsure about with state-of-the-art accuracy by dynamically re-weighting abstention rewards based on trajectory consistency during training.

Stop naively aggregating knowledge for KB-VQA: MaS-VQA's Mask-and-Select mechanism shows how to prune irrelevant image regions and knowledge fragments for better reasoning.