📝 Publications

Math-PUMA: Progressive Upward Multimodal Alignment to Enhance Mathematical Reasoning

Wenwen Zhuang*, Xin Huang*, Xiantao Zhang*, Jin Zeng

Project

• This paper introduces Math-PUMA, a novel training methodology that uses Progressive Upward Multimodal Alignment to significantly narrow the performance gap between textual and visual modalities in mathematical reasoning , achieving state-of-the-art results among open-source MLLMs.

The Escalator Problem: Identifying Implicit Motion Blindness in AI for Accessibility

Xiantao Zhang

• This paper identifies “Implicit Motion Blindness” in MLLMs, exemplified by their failure to determine an escalator’s direction , and argues this flaw—caused by the prevalent frame-sampling paradigm —necessitates a paradigm shift from semantic recognition to robust physical perception to build trustworthy assistive AI for the visually impaired.

Preservation vs. Fabrication: An Ethical Framework of Consent, Transparency, and Integrity for Posthumous AI Art

Xiantao Zhang

• This paper proposes an ethical framework for posthumous AI art, demanding the artist’s explicit consent, mandatory transparency, and an integrity rule that forbids AI from fabricating new works where no verifiable intent exists.