Omni-modal orchestration
Training-free coordination of modality experts through LLM routing, persistent memory, and full-duplex interaction.
M.S. student at MAC LabXiamen University
Interactive multimodal intelligence
Tianyu Xie / 谢天宇
I study how AI systems can listen, see, reason, act, and revise continuously across time.
Omni models · Agentic systems
Long-video reasoning · Efficient inference
“The task we must set for ourselves is not to feel secure, but to be able to tolerate insecurity.”
我们需要为自己设定的任务，不是拥有安全感，而是能够接受不安全感。
Erich Fromm
Current research
Building systems that stay present.
I am an M.S. student admitted in Fall 2025 at the MAC Lab, Xiamen University, advised by Prof. Xiawu Zheng.
Tianyu XieMAC Lab · Xiamen Universityteery@stu.xmu.edu.cn
Agentic systems and world models
Policies, memory, tools, and HCI-oriented evaluation for AI systems that must complete grounded tasks.
Long-video reasoning
Event-aware and semantic-boundary-aware frame selection for efficient long-form video understanding.
Efficient and adaptive inference
Speculative decoding, online drafting, and lightweight adaptation for deployable foundation models.
Selected publications
Work across interaction, perception, and inference.
View the full publication index
ICML 2026 · First author · CCF-A
Training-Free Multimodal Large Language Model Orchestration
A modular orchestration framework with explicit expert routing, cross-modal memory, and full-duplex streaming interaction.
Audio-visual benchmark · First author
SocialOmni: Benchmarking Audio-Visual Social Interactivity in Omni Models
A benchmark for who speaks, when to intervene, and how to generate natural interruptions in audio-visual interaction.
CVPR 2026 · CCF-A
Wavelet-based Frame Selection by Detecting Semantic Boundary for Long Video Understanding
Semantic-boundary-aware frame selection for compact, informative, and coherent long-video understanding.
News
Recent milestones.
Training-Free Multimodal Large Language Model Orchestration was accepted to ICML 2026 as first-author work.
SocialOmni was released on arXiv as a first-author benchmark for audio-visual social interactivity.
WFS-SB was accepted to CVPR 2026 for efficient long-video understanding.
Joined Xiamen University MAC Lab as a master's student in Artificial Intelligence.
Submitted routing-guided expert selection work for mitigating gradient interference in MoE models.