Tianyu Chen

White Sands, New Mexico

I am a third-year PhD student at the University of Texas at Austin, supervised by Professor Mingyuan Zhou. Before joining UT, I obtained my master’s degree in Statistics from the University of Chicago, where I closely collaborated with Kevin Bello, Bryon Aragam, Pradeep Ravikumar, Francesco Locatello, and supervised by Professor Jingshu Wang. 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 have broad research interests in statistical machine learning. Specifically, my research interests lie in:

Generative Models: diffusion models, model distillation, and controllable image editing.
Reinforcement Learning for LLM: post-training for both autoregressive (AR) and diffusion language models (dLLMs), focusing on reasoning enhancement and algorithm design.
Agentic AI – constructing autonomous agents for data generation and multi-turn evaluation.
Statistical Sampling & Inference – neural posterior sampling and hypothesis testing under uncertainty.

news

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 The Surprising Strength of Weak Classifiers (for validating neural posterior estimates).
Apr 01, 2026	Excited to start as a Student Researcher at Google DeepMind!
Jan 01, 2026	Two papers accepted at ICLR 2026: EdiVal-Agent (object-centric automated image editing evaluation, project page) and Score Distillation Beyond Acceleration (generative modeling from corrupted data).
Nov 01, 2025	Excited to announce that two papers are accepted by NeurIPS 2025! 🎉 CoLT (The Conditional Localization Test for assessing neural posterior estimates) is accepted as a Spotlight, and Improving Data Efficiency for LLM Reinforcement Fine-tuning introduces difficulty-targeted online data selection for more efficient LLM alignment. See you in San Diego!
Sep 01, 2025	Our new paper EdiVal-Agent introduces an object-centric VLM agent for automated, fine-grained evaluation of multi-turn image editing! Our benchmark reveals that SeedDream v4.0 > Nano Banana > GPT-Image-1 in multi-turn editing quality. The project received media coverage and is showcased on our project page.

selected publications

ICML 2026

REAL: Regression-Aware Reinforcement Learning for LLM-as-a-Judge

Yasi Zhang^*, Tianyu Chen^*, Mingyuan Zhou, and 3 more authors

The 43rd International Conference on Machine Learning. (Collaboration with Google DeepMind), 2026

Bib PDF

@article{chen2026real,
  title = {REAL: Regression-Aware Reinforcement Learning for LLM-as-a-Judge},
  author = {Zhang, Yasi and Chen, Tianyu and Zhou, Mingyuan and Leong, Oscar and Wu, Ying Nian and Lukasik, Michal},
  journal = {The 43rd International Conference on Machine Learning. (Collaboration with Google DeepMind)},
  year = {2026},
}

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

Bib PDF Code Website

@article{chen2026edival,
  title = {EdiVal-Agent: An Object-Centric Framework for Automated, Scalable, Fine-Grained Evaluation of Multi-Turn Editing},
  author = {Chen, Tianyu and Zhang, Yasi and Zhang, Zhi and Yu, Peiyu and Wang, Shu and Wang, Zhendong and Lin, Kevin and Wang, Xiaofei and Yang, Zhengyuan and Li, Linjie and Lin, Chung-Ching and Xie, Jianwen and Leong, Oscar and Wang, Lijuan and Wu, Ying Nian and Zhou, Mingyuan},
  journal = {The 14th International Conference on Learning Representations},
  year = {2026},
  media = {https://mp.weixin.qq.com/s/jlVDWl7wJ8SGFcGdRG7wGA},
  tldr = {An object-centric VLM agent that automates fine-grained, multi-turn image-editing evaluation, eliminating reliance on human annotators.}
}

ICLR 2026

Score Distillation Beyond Acceleration: Generative Modeling from Corrupted Data

Tianyu Chen^*, Yasi Zhang^*, Zhendong Wang, and 2 more authors

The 14th International Conference on Learning Representations, 2026

Bib PDF

@article{chen2026scoredistillation,
  title = {Score Distillation Beyond Acceleration: Generative Modeling from Corrupted Data},
  author = {Chen, Tianyu and Zhang, Yasi and Wang, Zhendong and Wu, Ying Nian and Zhou, Mingyuan},
  journal = {The 14th International Conference on Learning Representations},
  year = {2026},
  tldr = {Score distillation is not merely an acceleration technique — it enhances generation quality from corrupted data, both empirically and theoretically.}
}

NeurIPS 2025

CoLT: The conditional localization test for assessing the accuracy of neural posterior estimates

Tianyu Chen, Vansh Bansal, and James G. Scott

Advances in Neural Information Processing Systems 2025, 2025

Bib PDF Code

@article{chen2025colt,
  title = {CoLT: The conditional localization test for assessing the accuracy of neural posterior estimates},
  author = {Chen, Tianyu and Bansal, Vansh and Scott, James G.},
  journal = {Advances in Neural Information Processing Systems 2025},
  year = {2025},
  tldr = {A principled conditional localization test for validating neural posterior estimators. NeurIPS 2025 Spotlight.}
}

ICML 2026

The surprising strength of weak classifiers for validating neural posterior estimates

Vansh Bansal^*, Tianyu Chen^*, and James G. Scott

The 43rd International Conference on Machine Learning, 2026

Bib PDF

@article{chen2026weakclassifiers,
  title = {The surprising strength of weak classifiers for validating neural posterior estimates},
  author = {Bansal, Vansh and Chen, Tianyu and Scott, James G.},
  journal = {The 43rd International Conference on Machine Learning},
  year = {2026},
}

NeurIPS 2024

Diffusion Policies creating a Trust Region for Offline Reinforcement Learning

Tianyu Chen, Zhendong Wang, and Mingyuan Zhou

Advances in Neural Information Processing Systems 2024, 2024

Bib PDF Code

@article{chen2024diffusion,
  title = {Diffusion Policies creating a Trust Region for Offline Reinforcement Learning},
  author = {Chen, Tianyu and Wang, Zhendong and Zhou, Mingyuan},
  journal = {Advances in Neural Information Processing Systems 2024},
  year = {2024},
  url = {https://arxiv.org/abs/2405.19690},
  tldr = {Diffusion loss can be used to distill one-step policy and encourage mode-seeking.}
}

NeurIPS 2023

iSCAN: identifying causal mechanism shifts among nonlinear additive noise models

Tianyu Chen, Kevin Bello, Bryon Aragam, and 1 more author

Advances in Neural Information Processing Systems 2023, 2023

Bib PDF Code

@article{chen2024iscan,
  title = {iSCAN: identifying causal mechanism shifts among nonlinear additive noise models},
  author = {Chen, Tianyu and Bello, Kevin and Aragam, Bryon and Ravikumar, Pradeep},
  journal = {Advances in Neural Information Processing Systems 2023},
  volume = {36},
  year = {2023},
  url = {https://arxiv.org/abs/2306.17361},
  tldr = {Score matching help directly detect shifted nodes among different graphs.}
}

PNAS

Model-based trajectory inference for single-cell rna sequencing using deep learning with a mixture prior

Tianyu Chen^*, Jin-Hong Du^*, Ming Gao, and 1 more author

Proceedings of the National Academy of Sciences, 2024

Bib PDF Code

@article{du2020model,
  title = {Model-based trajectory inference for single-cell rna sequencing using deep learning with a mixture prior},
  author = {Chen, Tianyu and Du, Jin-Hong and Gao, Ming and Wang, Jingshu},
  journal = {Proceedings of the National Academy of Sciences},
  pages = {Vol. 121 | No. 37},
  year = {2024},
  url = {https://www.pnas.org/doi/10.1073/pnas.2316256121},
  tldr = {A VAE with a hierarchical prior offers a comprehensive pipeline for integrating multi-omic data, correcting batch effects, inferring pseudotime, and conducting differential analysis.}
}