Tianjian Li

Center for Language and Speech Processing, Johns Hopkins University

prof_pic.jpg

Hi👋, I’m Tianjian! I’m a PhD student in Computer Science at Johns Hopkins University, proudly advised by Prof. Daniel Khashabi. I am also a research intern at Meta FAIR, where I have the privilege to be advised by Tianlu Wang.

Previously, I completed my Master’s degree in Computer Science at JHU. I worked with my wonderful advisors Kenton Murray and Philipp Koehn. Before that, I was an undergraduate at New York University.

My research lies at the intersection between machine learning and natural language processing.

I prefer solutions that are simple, generalizable, and theoretically sound.

If you have anything to share with me, please feel free to contact me through my email: tli104 at jhu.edu

news

Mar 20, 2026 Our new work: Reasoning over mathematical objects: on-policy reward modeling and test time aggregation is out! In this work we 1) built and released training data for deriving mathematical objects; 2) show that on-policy RL with strong verifier boosts performance, and 3) on-policy training on parallel generation + verification further boosts the performance.
Mar 1, 2026 I will be returning to Meta AI Research (FAIR) at NYC as a research intern in summer 2026!
Sep 4, 2025 Our new work: Jointly Reinforcing Diversity and Quality in Language Model Generations is out! In this work, we studied how to make language models generate diverse outputs without sacrificing quality using online reinforcement learning.
May 1, 2025 SimpleMix is accepted to ICML 2025 !! In this work, we studied the interplay between on- and off-policy data in preference optimization.
Jan 23, 2025 3 papers are accepted to NAACL🎉, which includes my work on training on heavily imbalanced datasets, Jack’s work on making language models produce verbatim quotes from training data, and Yining’s work on evaluating the creativity of language models on code generation. I am super grateful to my wonderful co-authors!

selected publications

  1. Reasoning over mathematical objects: on-policy reward modeling and test time aggregation
    Pranjal Aggarwal, Marjan Ghazvininejad, Seungone Kim, and 18 more authors
    2026
    arXiv
  2. arxiv
    Jointly Reinforcing Diversity and Quality in Language Model Generations
    Tianjian Li, Yiming Zhang, Ping Yu, and 5 more authors
    2025
    arXiv
  3. ICML
    SIMPLEMIX: Frustratingly Simple Mixing of Off- and On-policy Data in Language Model Preference Learning
    Tianjian Li, and Daniel Khashabi
    In ICML 2025, 2025
    arXiv
  4. NAACL
    Upsample or Upweight? Balanced Training on Heavily Imbalanced Datasets
    Tianjian Li, Haoran Xu, Weiting Tan, and 2 more authors
    In NAACL 2025, 2025
    arXiv
  5. NAACL
    Benchmarking Language Model Creativity: A Case Study on Code Generation
    Yining Lu, Dixuan Wang, Tianjian Li, and 2 more authors
    In NAACL 2025, 2025
    arXiv
  6. NAACL
    Verifiable by Design: Aligning Language Models to Quote from Pre-Training Data
    Jingyu Zhang, Marc Marone, Tianjian Li, and 2 more authors
    In NAACL 2025, 2025
    arXiv
  7. ICLR
    Error Norm Truncation: Robust Training in the Presence of Data Noise for Text Generation Models
    Tianjian Li, Haoran Xu, Philipp Koehn, and 2 more authors
    In ICLR 2024
    (Spotlight - Top 5%), 2025
    arXiv
  8. ACL
    Why Does Zero-shot Cross-lingual Generation Fail? An Explaination and A Solution
    Tianjian Li, and Kenton Murray
    In ACL 2023 (Findings), 2025
    arXiv