Yunze Xiao

The paper proposes Information Sufficiency (IS) as a comprehensive framework for privacy-preserving LLM communication, demonstrating that free-text pseudonymization outperforms existing suppression and generalization methods, especially in multi-turn conversations.

The paper introduces Dr. DocBench, a difficulty-aware, comprehensive benchmark designed to rigorously test expert-level and challenging document parsing capabilities for VLMs, demonstrating that current state-of-the-art models fail on complex, domain-specific structures.

This paper introduces the ExpressionCueLens framework to analyze how anthropomorphism expressions are used in human-AI companion agent interactions on Reddit and XiaoHongShu, revealing cultural and platform differences.