Di Yang 杨迪

Ph.D. Researcher | Generative Models, World Models & Intelligent Agents

3D/4D Generation · Robotic Planning · Agent Learning · Visual Reasoning

🏛️ The College of William & Mary, Williamsburg, VA

🤝 Also working with: Dr. Haipeng Chen, Dr. Yufei Wang

✉️ Email: dyang06@wm.edu / diyang_uestc@outlook.com

I am a Ph.D. student in Data Science at The College of William & Mary, supervised by Dr. Yanhai Xiong, and also work with Dr. Haipeng Chen. Currently, doing research with SparcAI Inc. (mentored by Dr. Yufei Wang). Previously, I received my M.Sc. in Electrical Engineering from National University of Singapore (NUS) (advised by Dr. John S. Ho, and also interned with Dr. Xuejiao Zhao) and B.Eng. in Microelectronic Science and Engineering from University of Electronic Science and Technology of China (UESTC). Prior to my Ph.D., I worked as a Digital Logic Engineer at Cambricon Technologies on large-scale AI accelerator hardware verification.

My research interests lie at the intersection of Generative Modeling, World Models, and Intelligent Agents. My research spans 3D/4D generation, structured world representations, and learning-based planning, with growing interests in agent learning, multimodal reasoning, and decision-making. I am always open to research collaborations and discussions—feel free to reach out!

✉️ Email (WM) ✉️ Email (Outlook) 🎓 Google Scholar 📖 OpenReview 🐙 GitHub 📄 CV (PDF)

News & Updates

2026.10 📄 New preprint: "What Explicit Physics Buys a Manipulation World Model" is out!

2026.10 📄 New preprint: "Ask About the Right Moment: Training-Free Object Identity Reasoning for Video-Language Models" (TAIR) released on arXiv:2610.31471!

2026.10 💡 New blog post online: "Some Thoughts on Spatial Representation: From AI Chips to Embodied Intelligence".

2026.10 📝 New research blog post: "Does a Failed Agent’s Workspace Help the Next One? A Controlled Pilot on Failure-Conditioned Recovery Between Coding Agents", full open-source code and data on GitHub!

2026.09 🎉🎉 Double acceptance! Both "VTV-FM" and "HyCO" have been Accepted by NeurIPS 2026!

2026.07 🎉 Our paper "LENS: Learning to Navigate with Active Search for Partially Observable MAPF in Unknown Environments" was Accepted by Transactions on Machine Learning Research (TMLR)!

2026.06 Awarded the Summer Research Fellowship at The College of William & Mary.

2026.05 Developed Channel Match for high-resolution 3D distillation on TRELLIS.2.

2025.08 🎉 Our paper "UP-Bench: A Benchmark for Underwater Path Planning Algorithms" was accepted for an Oral Presentation at KDD 2025 (Benchmark Track)!

2025.06 Awarded the Summer Research Fellowship at The College of William & Mary.

2024.09 Joined The College of William & Mary as a Ph.D. student in Data Science.

Research Interests

3D/4D Generative Modeling & World Models

Explicit 4D spatiotemporal representations (Sparc4D), few-step 3D distillation (Channel Match), continuous-time flow matching (VTV-FM), and physical world action models (WAM).

Robotic Planning & Embodied Intelligence

Learning-based spatial planning under partial observability (LENS, TMLR), realistic 3D simulation platforms (UP-Bench, KDD Oral), and closed-loop robot manipulation policies (WAM).

Agent Learning & Multimodal Reasoning

Investigating multi-agent coordination, failure-conditioned recovery in coding agents, and visual-spatial reasoning in multimodal foundation models for complex decision-making.

Selected Publications

Preprints & Under Review

What Explicit Physics Buys a Manipulation World Model

Di Yang, Muyang Liu, Junzhou Chen, Yufei Wang, Yanhai Xiong

Preprint, 2026 Preprint

Investigates when explicit representations of unobservable physical properties (e.g., mobility, friction, mass distribution) benefit manipulation world models observing interaction history. Across six simulated manipulation environments and two model families (PointWorld and DINO-WM), we reveal that explicit factors reduce wasted pushing on locked targets by 37% primarily through geometric output constraints rather than dynamics conditioning. We are currently working on real-world robotic arm deployment and validation.

[arXiv (Coming Soon)] [PDF] [Code]

@article{yang2026what,
  title={What Explicit Physics Buys a Manipulation World Model},
  author={Yang, Di and Liu, Muyang and Chen, Junzhou and Wang, Yufei and Xiong, Yanhai},
  journal={arXiv preprint},
  year={2026}
}

Ask About the Right Moment: Training-Free Object Identity Reasoning for Video-Language Models

Di Yang, Xiaochang Li, Muyang Liu, Yufei Wang, Yanhai Xiong

arXiv:2610.31471, 2026 Preprint

Introduces Track, Then Bind (TAIR), a training-free framework for persistent visual object identity in Video-LMs. By decoupling semantic identity binding at the evidence moment from temporal trajectory tracking, frozen models solve VET-Bench (100%) and achieve state-of-the-art results on Perception Test and TOC-Bench. We are actively exploring its downstream applications across object-centric video search, persistent visual memory for video agents, and occlusion-aware robotic manipulation.

[arXiv] [PDF] [Demo] [Code]

@article{yang2026ask,
  title={Ask About the Right Moment: Training-Free Object Identity Reasoning for Video-Language Models},
  author={Yang, Di and Li, Xiaochang and Liu, Muyang and Wang, Yufei and Xiong, Yanhai},
  journal={arXiv preprint arXiv:2610.31471},
  year={2026}
}

Sparc4D: A Compact Explicit 4D Representation for Dynamic Scenes

Di Yang, Zhihao Li, Yanhai Xiong, Yufei Wang

arXiv:2610.01229, 2026 Preprint

A feed-forward autoencoder encoding monocular videos into a compact, sparse 4D scene state by sharing static features and compressing dynamic motion into spatially anchored temporal slots. Decodes directly into 2D Gaussian surfels, achieving competitive novel-view synthesis (21.97 dB on MultiCamVideo) at ~0.66M state size (~11× smaller than single-frame baselines) with zero-shot transfer to real DyCheck and Neu3D videos.

Note: The [PDF] link above has been updated to our latest refined manuscript (featuring robot manipulation downstream results on StackCube). The updated version will be synced to arXiv soon!

[arXiv] [PDF] [Project Page] [Code (Coming Soon)]

@article{yang2026sparc4d,
  title={Sparc4D: A Compact Explicit 4D Representation for Dynamic Scenes},
  author={Yang, Di and Li, Zhihao and Xiong, Yanhai and Wang, Yufei},
  journal={arXiv preprint arXiv:2610.01229},
  year={2026}
}

Channel Match: Calibrating Shape Latents for Few-Step 3D Generation

Di Yang, Z. Li, Y. Xiong, Y. Wang

Still under review Under Review

Distills a native 3D generator into a two-step student via continuous self-consistency training with a custom JVP FlashAttention kernel. Channel Match calibrates sampling-induced latent distribution drift, achieving 4.5× faster geometry sampling on TRELLIS.2 (1024³ resolution) without texture artifacts.

[PDF] [Code (Coming Soon)]

@article{yang2026channelmatch,
  title={Channel Match: Calibrating Shape Latents for Few-Step 3D Generation},
  author={Yang, Di and Li, Z. and Xiong, Y. and Wang, Y.},
  journal={Under Review},
  year={2026}
}

Peer-Reviewed

HyCO: A Hybrid Neural Solver for Combinatorial Optimization

Y. Li, Di Yang*, H. Chen, Y. Xiong (*Co-first author)

Annual Conference on Neural Information Processing Systems (NeurIPS), 2026 NeurIPS 2026

Designs an RL-to-diffusion handover using policy entropy and model disagreement to combine sequential construction with generative solution completion, reducing TSP-1000 optimality gap from 15.4% to 4.6%.

[OpenReview] [arXiv] [PDF] [Code]

@inproceedings{li2026hyco,
  title={HyCO: A Hybrid Neural Solver for Combinatorial Optimization},
  author={Li, Y. and Yang, Di and Chen, H. and Xiong, Y.},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2026}
}

VTV-FM: Flow Matching through Variational Terminal-Velocity Closure

H. Jiang, Y. Li, Di Yang*, Y. Xiong, H. Chen, Y. He (*Co-first author)

Annual Conference on Neural Information Processing Systems (NeurIPS), 2026 NeurIPS 2026

Co-developed phase-space flow matching with cubic Hermite trajectories and a closed-form minimum-acceleration variational closure, achieving FID 2.99 on CIFAR-10 versus 3.45-3.73 for baselines.

[OpenReview] [PDF] [Code]

@inproceedings{jiang2026vtvfm,
  title={VTV-FM: Flow Matching through Variational Terminal-Velocity Closure},
  author={Jiang, H. and Li, Y. and Yang, Di and Xiong, Y. and Chen, H. and He, Y.},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2026}
}

LENS: Learning to Navigate with Active Search for Partially Observable MAPF in Unknown Environments

Di Yang et al.

Transactions on Machine Learning Research (TMLR), 2026 Accepted

A hybrid Multi-Agent Pathfinding architecture decoupling macro-topological spatial reasoning (via U-Nets) from micro-level collision avoidance, achieving generalization comparable to 85M foundation models with <0.2% training data.

[OpenReview] [PDF] [Code]

@article{yang2026lens,
  title={LENS: Learning to Navigate with Active Search for Partially Observable MAPF in Unknown Environments},
  author={Yang, Di and others},
  journal={Transactions on Machine Learning Research},
  year={2026}
}

UP-Bench: A Benchmark for Underwater Path Planning Algorithms

Di Yang, Yanhai Xiong

ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), 2025 Benchmark Track Oral

A configurable open-source 3D simulation benchmark platform incorporating realistic fluid dynamics, partial observability, and dynamic obstacles, benchmarking Deep RL (SAC) and classical planning algorithms.

[ACM DL / Paper] [PDF] [Code]

@inproceedings{yang2025upbench,
  title={UP-Bench: A Benchmark for Underwater Path Planning Algorithms},
  author={Yang, Di and Xiong, Yanhai},
  booktitle={Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  year={2025}
}

Selected Research & Industry Experience

Research Collaboration & Research Intern Jan 2026 – Present

SparcAI Inc. (Remote) • Advisor: Dr. Yufei Wang

Large-Scale 3D Data: Contributed to curating and processing a ~1M-scale 3D asset dataset, spanning quality filtering, geometric repair, and multi-view data generation for 3D generative modeling.
3D Generation & Acceleration: Developed Channel Match, distilling a native 3D generator to 2 sampling steps (4.5× faster); implemented a custom Triton JVP FlashAttention kernel.
4D World Representation (Sparc4D): Developed a compact dynamic-scene representation with ~12× smaller state; matched StackCube predictive-planning performance using ~1,170 tokens.
VLA & Robotic Manipulation: Integrated robot-centric spatial representations into VLA policies (xiaomirobotics-1, π0.5); built OmniGibson/BEHAVIOR pipelines with failure-case augmentation and validated manipulation on a physical SO-101 robot.
Next-Gen World Models (Ongoing): Exploring unified physical world-state tokenization and compact 4D representations for scalable world modeling and embodied intelligence.

Ph.D. Researcher Sep 2024 – Present

The College of William & Mary • Advisor: Dr. Yanhai Xiong

HyCO: Designed RL-to-diffusion handover for combinatorial optimization; reduced TSP-1000 optimality gap from 15.4% to 4.6% (NeurIPS 2026, arXiv:2609.07990).
LENS & UP-Bench: Proposed LENS (TMLR) for partially observable MAPF; designed UP-Bench (KDD 2025 Oral) 3D underwater simulation benchmark.
VTV-FM: Co-developed second-order phase-space flow matching with variational terminal-velocity closure (NeurIPS 2026).

Digital Logic Engineer Aug 2022 – Jul 2024

Cambricon Technologies Co., Ltd. (Beijing, China)

Built functional verification pipelines for large-scale AI accelerator chips; owned system-level performance verification for low-latency data flow.

Graduate Researcher Jan 2021 – Jan 2022

National University of Singapore (NUS) • Advisor: Dr. John S. Ho

Conducted graduate research in wireless micro-systems and intelligent embedded hardware.