邓安东 Andong Deng

I am a forth-year PhD student at the Institute of Artificial Intelligence (IAI), University of Central Florida, advised by Dr. Chen Chen.

From 2021 to 2022, I spent a wonderful period of time at GeWu Lab, working with Dr. Di Hu . Prior to that, I obtained my master's degree from Shanghai Jiao Tong University in 2021, working with Dr. Weiwei Cai. I obtained my bachelor's degree from Sichuan University.

Beyond research, I regularly spend time on photography (see the gallery below), playing hold'em (low stakes smasher), music (Erhu, Level 10 certification), and a variety of sports (e.g., basketball, skateboarding).

Email  /  Curriculum Vitae  /  X (Twitter)  /  知乎 Zhihu  /  Google Scholar

My research focuses on multimodal AI—how we can build models that perceive, ground, and reason over long, complex procedures with different modalities involved. I develop benchmarks that stress advanced reasoning, grounding frameworks that align language with space-time, and representation learning methods that scale robustly.

• Video Understanding / Reasoning Benchmarks: VEBEnch(CVPR Findings 2026), SciVideoBench(ICML 2026), GroundMoRe(CVPR 2025), Sports-QA(IJCV 2026)
  Rigorous evaluations of multimodal reasoning in scientific and real-world domains.
• Video Spatial / Temporal Grounding Frameworks: Seq2Time(CVPR 2025), MoRA(CVPR 2025)
  Mapping language to precise space-time segments with sequence-level localization.
• Representation Learning: BEAR(ICCV 2023), Inadequate Pre-training(ICCV 2023), OGM-GE(CVPR 2022), MoPRL(TCSVT 2022)
  Foundation features and training recipes for robust visual perception and reasoning.

^* equal contribution

We introduce SciVideoBench, the first scientific video reasoning benchmark!

Video LLMs must not ignore the pixel dynamics in plain sight: models and benchmarks must make spatiotemporal evidence unavoidable, and must fail when motion and state transitions are ignored or contradicted.

We introduce VEBench, the first comprehensive benchmark designed to evaluate both editing knowledge understanding and operational reasoning in realistic video editing scenarios

We introduce AIR, a training-free approach for Adaptive, Iterative, and Reasoning-based frame selection!

We present Seq2Time, a data-oriented method to enhance the time perception ability of video LLMs.

We present GroundMoRe, a new benchmark for novel Motion-Grounded Video Reasoning; and MoRA, a video LLM that is able to reason and perceive motion at the pixel level.

In this paper, we introduce Sports-QA, the first dataset specifically designed for the sports-related VideoQA task.

We propose OIS: order-aware interactive segmentation, to explicitly integrate missing relative depth information into 2D interactive segmentation.

We introduce BEAR, a new BEnchmark on video Action Recognition. BEAR is a collection of 18 video datasets grouped into 5 categories, which covers a diverse set of real-world applications.

Inadequately pre-trained models often extract better features than fully pre-trained ones. And deep models tend to first learn spectral components corresponding to large singular values.

A motion prior distribution is utilized to construct the pose representation. MoPRL achieves the state-of-the-art performance by an average improvement of 4.7% AUC on several challenging datasets.

We propose a Modality Noise Filter module to erase the irrelevant noise in teacher modality with cross-modal context and design a Contrastive Semantic Calibration module to adaptively distill useful knowledge for target modality, by referring to the differentiated sample-wise semantic correlation in a contrastive fashion.

Modulate gradients of two modalities to adaptively balance the optimization process of multimodal learning.

Conference Reviewer: CVPR, ECCV, ICCV, ICML, ICLR, NeurIPS

Journal Reviewer: IEEE IoT, IEEE TNNLS, IJCV, IEEE TAPMI

Updated: June 30th, 2026.

Thanks Jon Barron for this amazing template.