CV
General Information
| Full Name | Tianyu Chen |
| tianyuchen@utexas.edu | |
| Location | Austin, TX |
| Expected Graduation | May 2028 |
| Languages | English, Chinese |
| Links | Website · Google Scholar · GitHub · LinkedIn |
Industry Experience
-
03/2026 – now Student Researcher
Google DeepMind | Gemini - Mentor: Jiaxing Wu
- Post-training the Gemini Embedding model for agentic search, turning it into an autoregressive retriever that iteratively issues queries and performs multi-hop visual retrieval over million-scale corpora that are far too large to fit into the model's context window.
- Designing the reinforcement learning (RL) algorithm that trains this retriever end-to-end from multi-hop retrieval outcomes, improving instruction following on compositional image retrieval tasks.
-
02/2025 – 03/2026 Research Scientist Intern
Microsoft Research | MAI Superintelligence - Mentors: Zhendong Wang, Lijuan Wang
- Developed an object-centric vision-language model (VLM) agent that automates fine-grained, multi-turn image-editing evaluation, reducing reliance on human annotators for benchmark construction. The project received media coverage in prominent Chinese technology outlets and is showcased on the project website.
- Led large language model (LLM) post-training research on mathematical reasoning, LLM-as-a-judge alignment, and on-policy distillation, including training-recipe development for autoregressive and diffusion-based language models.
Education
-
2023-now PhD in Statistics & Data Sciences
University of Texas at Austin, Austin, United States - GPAs 4.0/4.0
-
2021-2023 Master in Statistics
University of Chicago, Chicago, United States - GPAs 3.97/4.0
-
2017-2021 Bachelor in Statistics
Fudan University, Shanghai, China - GPAs 3.7/4.0
Honors & Awards
-
2026 Intern and Student Researcher Award
Google DeepMind -
2025 Spotlight Paper (CoLT)
NeurIPS 2025
Academic Interests
-
RL for LLM post-training (autoregressive and diffusion LLMs, LLM-as-a-Judge, data-efficient GRPO)
-
Reward design and agentic evaluation (VLM agents, multi-turn image editing)
-
Agentic search, multi-hop retrieval, and test-time adaptation of embedding models
-
Distillation beyond acceleration (one-step policies, generative modeling from corrupted data, causal estimation)
-
Statistical foundations for reliable ML (neural posterior estimation, two-sample tests, causal mechanism shifts)
Other Interests
- Hobbies: Sleeping, Tennis, Climbing etc.