Tianyu Chen

University of Texas at Austin; tianyuchen@utexas.edu

prof_pic.jpg

White Sands, New Mexico

I am a fourth-year PhD candidate in Statistics at the University of Texas at Austin (expected graduation: May 2028), advised by Professor Mingyuan Zhou, working on reinforcement learning for LLM post-training, agentic search, and automated evaluation. I am currently a Student Researcher at Google DeepMind (Gemini), where I received the 2026 Intern and Student Researcher Award, and was previously a Research Scientist Intern at Microsoft Research (MAI Superintelligence).

Resume · Google Scholar · GitHub · LinkedIn · Email

Before UT, I obtained my master’s degree in Statistics from the University of Chicago, supervised by Professor Jingshu Wang and closely collaborating with Kevin Bello, Bryon Aragam, Pradeep Ravikumar, and Francesco Locatello. I earned my Bachelor’s degree in Statistics from Fudan University, where I spent some of the most memorable moments of my life.

Research Interests: I build reinforcement learning and distillation methods that make LLMs and multimodal agents reason, search, and evaluate reliably, grounded in statistical rigor.

  • RL for LLM post-training: algorithms for reasoning and alignment in both autoregressive and diffusion LLMs, including regression-aware policy gradients for LLM-as-a-Judge (REAL, ICML 2026) and data-efficient GRPO via difficulty-targeted selection and rollout replay (NeurIPS 2025).
  • Reward design & agentic evaluation: turning LLM/VLM judges, detectors, and rankers into trustworthy reward signals; e.g., EdiVal-Agent (ICLR 2026), an object-centric agent that autonomously synthesizes instructions and verifies multi-turn edits, and MT-EditFlow, multi-turn RL with reward aggregation that resists reward hacking.
  • Agentic search & test-time adaptation: RL-trained autoregressive retrievers for multi-hop retrieval over million-scale corpora (Google DeepMind), and reusing ranking rewards as test-time state for frozen embedding models (TTT-Embed).
  • Distillation beyond acceleration: score and policy distillation as a mechanism to improve models, from one-step policies for offline RL (DTQL, NeurIPS 2024) to generative modeling from corrupted data (ICLR 2026) and causal estimation (IWDD).
  • Statistical foundations for reliable ML: finite-sample tests for validating neural posterior estimators and generative models (CoLT, NeurIPS 2025 Spotlight; Conformal C2ST, ICML 2026) and identifiability of causal mechanism shifts (iSCAN, NeurIPS 2023).

news

Aug 12, 2026 New preprint from my work at Google DeepMind: Test-Time Optimization of Query Embeddings with Ranking Aware Reward Maximization, a step toward agentic, multi-hop visual retrieval over million-scale corpora.
Jul 20, 2026 Honored to receive the 2026 Intern and Student Researcher Award at Google DeepMind! 🏆
Jun 01, 2026 New preprint MT-EditFlow: reinforcement learning for multi-turn image editing with flow matching.
Apr 30, 2026 Two papers accepted at ICML 2026! 🎉 REAL (Regression-Aware Reinforcement Learning for LLM-as-a-Judge, in collaboration with Google DeepMind) and Conformal C2ST (turning weak classifiers into strong two-sample tests for validating neural posterior estimates).
Apr 01, 2026 Excited to start as a Student Researcher at Google DeepMind!

selected publications

  1. Preprint
    Test-Time Optimization of Query Embeddings with Ranking Aware Reward Maximization
    Tianyu Chen, and Jiaxing Wu
    arXiv preprint arXiv:2608.12569. (Collaboration with Google DeepMind), 2026
    Distills ranking rewards into a learned vector in a frozen retriever’s embedding space: up to +8.36 nDCG@10 across 15 MTEB retrieval tasks, generalizing to unseen queries and tasks with no weight updates.
  2. Preprint
    MT-EditFlow: Reinforcement Learning for Multi-Turn Image Editing with Flow Matching
    Jiahui Huang, Yasi Zhang, Tianyu Chen, and 6 more authors
    arXiv preprint arXiv:2606.01985, 2026
    Flow-matching RL for multi-turn image editing with multi-reward advantage broadcasting: +6.85 turn-3 overall score on FLUX.1-Kontext-dev, surpassing Qwen-Image-Edit.
  3. ICML 2026
    REAL: Regression-Aware Reinforcement Learning for LLM-as-a-Judge
    Tianyu Chen*, Yasi Zhang*, Mingyuan Zhou, and 3 more authors
    The 43rd International Conference on Machine Learning. (Collaboration with Google DeepMind), 2026
    Regression-aware policy gradients for LLM-as-a-Judge: +8.40 Pearson and +7.20 Spearman over the SFT baseline on Qwen3-32B, with stronger out-of-domain generalization.
  4. ICLR 2026
    EdiVal-Agent: An Object-Centric Framework for Automated, Scalable, Fine-Grained Evaluation of Multi-Turn Editing
    Tianyu Chen*, Yasi Zhang*, Zhi Zhang, and 13 more authors
    The 14th International Conference on Learning Representations, 2026
    An object-centric VLM agent that automates fine-grained, multi-turn image-editing evaluation without human annotators; EdiVal-Bench covers 9 instruction types and 16 editing models.
  5. NeurIPS 2025
    Improving Data Efficiency for LLM Reinforcement Fine-tuning Through Difficulty-targeted Online Data Selection and Rollout Replay
    Yifan Sun, Jingyan Shen, Yibin Wang, and 4 more authors
    Advances in Neural Information Processing Systems 2025, 2025
    Difficulty-targeted online data selection and rollout replay cut GRPO fine-tuning time by 23-62% while matching the original performance.