About

I am a fourth-year PhD student at Renmin University of China (RUC) under the supervision of Professor Qin Jin at AIM3 Lab. Previously, I obtained my bachelor's degree from School of Computer Science and Engineering, University of Electronic Science and Technology of China (UESTC).

Email  /  WeChat  /  Github  /  Google Scholar

Research

My research interests center around computer vision, vision-and-language, and LLM/VLM agents. In particular, I work on:

• Multimodal agents. More recently, I have been exploring general-purpose agents, systems that can perceive, reason, and act across open-ended tasks and environments. Besides, I focus on visual content understanding and generation that is programmable and expressed as code, aiming to find representations that are more token-efficient for LLMs, that enhance models' foundational capabilities, and that are more editable for productivity scenarios.
• Video understanding. Video is the continuous two-dimensional projection of the four-dimensional spatio-temporal world. Correspondingly, I have studied video from three perspectives: performing complex visual reasoning and content understanding over its rich semantics; learning knowledge of the world and 3D representations from its visual content; and building long-context model architectures that are both fast and accurate.
• Human-environment interaction. Human-environment interaction embodies rich knowledge of interaction, motion, and physics. I have focused on learning generalizable representations from first-person (egocentric) video, with two goals in mind: empowering VR/AR glasses with egocentric understanding to serve as an everyday AI assistant, and building robots with human-like capabilities. I have also briefly studied 3D human-object interaction modeling: building image-to-3D models that produce explicit 3D representations, which can support robot data construction and real-world operation in VR/AR.

Multimodal Large Language Models

We introduce TimeViper, a hybrid vision-language model for long video understanding. This work represents an initial step towards developing, interpreting, and compressing hybrid Mamba-Transformer architectures.

We introduce Time-R1 framework, TimeRFT training, and TVGBench for LVLM evaluation, to advance the field of using LVLM for temporal video grounding. Time-R1 achieves state-of-the-art performance on TVG using only 2.5K data for RL fine-tuning, with improved performance on 4 VQA benchmarks.

We propose TimeZero, a reasoning-guided LVLM for temporal video grounding that extends inference through reinforcement learning to reason about video-language relationships. Achieves state-of-the-art performance on Charades-STA benchmark.

• EventLens: Event-Structure Reinforcement Learning for Video Understanding, Guowen Zhang, Boshen Xu, Zihao Yue, Ziheng Wang, Xiaokun Liu, Xin Tao, Wenyu Qin, Pengfei Wan, Qin Jin^†. NeurIPS, 2026.
• Depth as Verifiable Geometric Feedback: Enhancing Spatial Reasoning in Multimodal Large Language Models, Yepeng Zhang, Ruohang Xu, Zihao Yue, Boshen Xu, Qin Jin^†. EMNLP Findings, 2026.
• Video-OPD: Efficient Post-Training of Multimodal Large Language Models for Temporal Video Grounding via On-Policy Distillation, Jiaze Li, Hao Yin, Haoran Xu, Boshen Xu, Wenhui Tan, Zewen He, Jianzhong Ju, Zhenbo Luo, Jian Luan. ICML, 2026. arxiv
• Revisor: Beyond Textual Reflection, Towards Multimodal Introspective Reasoning in Long-Form Video Understanding, Jiaze Li, Hao Yin, Wenhui Tan, Jingyang Chen, Boshen Xu, Yuxun Qu, Yijing Chen, Jianzhong Ju, Zhenbo Luo, Jian Luan. CVPR, 2026. paper
• Unveiling Visual Biases in Audio-Visual Localization Benchmarks, Liangyu Chen, Zihao Yue, Boshen Xu, Qin Jin^†. ECCV AVGenL Workshop, 2024. arxiv
• StreamPro: From Reactive Perception to Proactive Decision-Making in Streaming Video, Ao Li, Zihan Xiao, Zihao Yue, Boshen Xu, Linli Yao, Jiaze Li, Pei Fu, Jianzhong Ju, Jian Luan, Qin Jin^†. arXiv, 2026. arxiv
• Xiaomi MiMo-VL-Miloco Technical Report, Jiaze Li, Jingyang Chen, Yuxun Qu, Shijie Xu, Zhenru Lin, Junyou Zhu, Boshen Xu, Wenhui Tan, Pei Fu, Jianzhong Ju, Zhenbo Luo, Jian Luan. arXiv, 2025. arxiv  /  code

Human-Object Interaction Understanding

We introduce EgoDTM, an Egocentric Depth- and Text-aware Model that bridges the gap between 2D visual understanding and 3D spatial awareness.

We propose EgoNCE++, an asymmetric contrastive learning pretraining objective to solve the EgoVLMs' weakness in distinguishing HOI combinations with word variations.

We propose POV, a view adaptation framework that enables transfer learning from multi-view third-person videos to egocentric videos.

We introduce OpenCat, a language modeling framework that reformulates HOI prediction as sequence generation.

3D Vision

We present SPAFormer, a transformer-based framework that leverages assembly sequences constraints with three part encodings to address the combinatorial explosion challenge in 3D-PA task.

Blog

Experience

Awards

Services