About Me
Hello! I’m Haotian (Eric) Xia (夏浩天), a second-year Ph.D. student in Computer Science at Rice University, advised by Prof. Hanjie Chen in the Chili Lab. I work on multimodal understanding and reasoning: building vision-language models that find the evidence that matters in images and video, then reason over it accurately and efficiently.
I earned my M.S. in Statistics at the University of California, Irvine, advised by Prof. Weining Shen, and my B.S. in Computer Science with Distinction at the University of California, Santa Barbara, advised by Prof. Yuan-Fang Wang and Prof. Linda Petzold.
Research
Multimodal understanding
Grounding language in visual evidence across images and long or streaming video.
Multimodal reasoning
Latent and adaptive reasoning, post-trained with reinforcement learning, for both accuracy and inference efficiency.
AI for sports and beyond
Benchmarks that test rules, tactics and physical decision-making, with sports as a demanding real-world testbed.
SPORTU · SportD · VisFinEval
News
- Oct 2026Received the Ken Kennedy Institute Andrew Ladd Memorial Excellence in Computer Science Graduate Fellowship (2026/27).
- Aug 2026SportD accepted as an Oral at the NeurIPS 2026 Workshop on Physical Understanding.
- Aug 2026New preprint: Dynamic Latent Reasoning (DyLaR) for video understanding and question answering.
- Apr 2026SportR published at ICLR 2026 in Rio de Janeiro.
- Nov 2025New preprint: DeepSport. VisFinEval published at EMNLP 2025 in Suzhou.
- Aug 2025Started my Ph.D. in Computer Science at Rice University.
- Apr 2025SPORTU published at ICLR 2025; FinEval published at NAACL 2025.
Selected Publications
* Equal contribution. All publications
- Perception Before Reasoning: Dynamic Latent Reasoning for Video Understanding and Question AnsweringDyLaR grounds a question in a short set of perception latents, then decides whether reasoning latents are needed. It improves accuracy on video benchmarks while generating fewer than 20 tokens per query.
- An end-to-end multimodal model that selects frames as it reasons, trained with a sports curriculum and then agentic reinforcement learning.
- A multi-sport image and video benchmark with human-authored chain-of-thought annotations and bounding boxes, testing reasoning from infraction identification to penalty prediction.
- A text and slow-motion video benchmark for multi-level sports reasoning, from basic rules to hard, knowledge-dependent video tasks.
- A benchmark of over 70,000 multiple-choice sports questions at three difficulty levels, from factual recall to complex scenario analysis, where language models still lag behind humans.
Beyond Research
I started playing volleyball at the age of seven and later competed as an NCAA Division I student-athlete on an athletic scholarship with the nationally top-3-ranked UCSB Men's Volleyball team, winning the 2021 Big West Conference title. I am also recognized as a National Level Athlete (国家一级运动员) in China.
If you'd like to talk about multimodal reasoning or sports analytics, feel free to email me.
