๐Ÿ‘จโ€๐ŸŽ“ About Me

I am a master's student in Information and Communication Engineering at the University of Science and Technology of China. Before joining USTC, I received a bachelor's degree in Data Science and Big Data Technology from Harbin Institute of Technology, Shenzhen.

My research focuses on consistency in camera-controlled and multi-shot video generation: how models preserve character identities and keep event states evolving under camera interaction and shot transitions. I also work on efficient visual representation, multimodal fusion, and knowledge distillation.

At the Beijing Innovation Center of Humanoid Robotics, I worked on camera-conditioned video generation and evaluation, proposed WRBench as first author, and explored action-conditioned world modeling and few-step action prediction.

My publications include first-author papers at ICLR 2026 and MICCAI 2024, and collaborative work at EMNLP 2026 and in Medical Image Analysis 2025.

I am looking for a full-time internship in a core foundation-model team, with interests in pre-training, RL post-training, and research. I welcome research collaborations. Please feel free to view my ็ฎ€ๅކ / CV or email me.

  • Camera Control & Multi-Shot Consistency I study how video models preserve identities and update event states as the camera moves and shots change.
  • Embodied AI I explore action-conditioned world models and few-step action prediction, and contribute to VLA analysis.
  • Efficient Visual Representation I develop efficient 3D perception through multimodal fusion and knowledge distillation.

๐Ÿ”ฅ News

๐Ÿ’ผ Experience

2025.11 - 2026.10

Beijing Innovation Center of Humanoid Robotics (X-Humanoid)

Large Model Algorithm Intern, Embodied World Model Team, Beijing.

  • Independently led camera-controlled video generation, covering pose conditioning, multi-source data pipelines, training, and evaluation. Diagnosed reliance on camera-motion text and failures to preserve object dynamics. Proposed WRBench as first author, delivering benchmark design, automatic scoring, human pairwise evaluation, and the paper.
  • Explored few-step action-policy sampling and fixed an initialization defect that disabled action conditioning; completed preliminary controlled checks.
  • Contributed to the core design and refinement of VLA-Trace, accepted to EMNLP 2026, and handled data processing for Pelican-Unify 1.0 to support joint generation of future video and actions.

๐Ÿ“˜ Selected Publications

Full list: Google Scholar Google Scholar citations: 192 h-index 6 ยท i10-index 5

๐ŸŒ Video World Models & Generative Video Evaluation

arXiv 2026 WRBench source figure showing viewpoint changes and state-consistency evaluation
First Author 2026.06

Current World Models Lack a Persistent State Core

J. Lu, D. Zhu, H. Shi, L. Cai, G. Tang, Y. Chen, J. Cao, D. Tang, Y. Zhang, Y. Dai, X. Ju.

  • Question: Is rendering decoupled from world-state evolution when camera motion changes only the observation?
  • Approach: WRBench separates camera execution, visual integrity, visible evolution, re-observation support, and conditional endpoint consistency across four native camera-control interfaces.
  • Finding: Across 9,600 videos from 23 models, viewpoint changes disrupted event evolution; better visuals and re-observation support did not reliably yield correct event endpoints.

๐Ÿค– Embodied AI & VLA Analysis

1 / 2
EMNLP 2026 VLA-Trace source figure showing representation and behavior tracing framework
Co-author CCF B 2026.08

VLA-Trace: Diagnosing Vision-Language-Action Models through Representation and Behavior Tracing

  • Question: How do internal VLA representations translate into decisions and rollout behavior?
  • Approach: VLA-Trace combines representation analysis, causal attention interventions, and behavioral probes across ฯ€0.5 and OpenVLA.
  • Finding: The study identifies distinct adaptation and action-routing patterns, while revealing limits in fine-grained semantic following.
arXiv 2026 Pelican-Unify source figure showing embodied intelligence model components
Co-author 2026.05

Pelican-Unify 1.0: A Unified Embodied Intelligence Model

  • Question: Can one embodied model connect visual understanding, reasoning, future prediction, and action generation?
  • Approach: Pelican-Unify uses a shared vision-language model, then jointly generates future video and action.
  • Training: Language, video, and action objectives are optimized together instead of as isolated specialist capabilities.

๐Ÿฅ Medical AI & Efficient 3D Perception

1 / 3
ICLR 2026 VeloxSeg overview source figure
First Author CCF A 2026.01

Johnson-Lindenstrauss Lemma Guided Network for Efficient 3D Medical Segmentation

J. Lu, L. Cai, Y. Chen, G. Tang, S. Jiang, H. Shi, Z. Xiong.

  • Task: Preserve 3D representation quality and fine detail under a limited computational budget.
  • Method: Combine JL-guided convolution and multiscale multimodal attention; transfer reconstruction texture priors through Gram matrices to avoid spatial alignment conflicts across tasks.
  • Result: Distillation improves AutoPET-II Dice by 2.80 points with the same 1.66M-parameter architecture and no additional inference cost.
arXiv 2025 DINOv3 PET/CT feature source figure
Co-first Author 2025.10

Does DINOv3 Set a New Medical Vision Standard?

C. Liu*, Y. Chen*, H. Shi*, J. Lu*, B. Jian*, J. Pan*, L. Cai*, et al.

  • Task: Test whether natural-image pretraining provides a general visual representation for medical tasks.
  • Method: Adapt frozen DINOv3 features to 2D/3D classification and 3D segmentation, comparing model scales and input resolutions.
  • Result: Surpasses medical-specific models on some tasks, but scaling gains vary and transfer remains limited in pathology, electron microscopy, and PET.
MICCAI 2024 H2ASeg source figure showing PET/CT hierarchical adaptive interaction architecture
First Author CCF-B, top 11% 2024.03

H2ASeg: Hierarchical Adaptive Interaction and Weighting Network for Tumor Segmentation in PET/CT Images

J. Lu, J. Chen, L. Cai, S. Jiang, Y. Zhang.

  • Task: Overcome limited information exchange between independent PET and CT encoders to exploit complementary semantic and structural cues.
  • Method: Use local and global bidirectional attention for cross-modal interaction, then adaptively select modality features by target relevance.
  • Result: Outperforms compared methods on AutoPET-II and Hecktor2022; ablations and weight analysis support hierarchical interaction and adaptive feature selection.

Additional Medical AI Publications

๐Ÿงฐ Open-source Work

I build practical tools for agent and research workflows.

Three projects per page.

1 / 2

๐ŸŽ“ Education

2025.09 - 2028.06

University of Science and Technology of China

M.S. in Information and Communication Engineering. Expected graduation: June 2028.

๐Ÿ“ Academic Service

Reviewer for ICLR 2027, NeurIPS 2027, and IEEE TNNLS.

๐Ÿ‘ฃ Site Visitors

Approximate visitor map and page-view counter.

Approximate Location

Approximate location: loading...

Location is estimated from public IP geolocation and may be imprecise; this page does not display your raw IP address.

Visitor map and counter powered by Flag Counter