Sining Zhoubian

About

I am a master's student in the Department of Computer Science and Technology at Tsinghua University, advised by Prof. Jie Tang in the Knowledge Engineering Group (KEG). I also intern at ZhipuAI. My work concerns how large language models reason and remember: search under sparse feedback, self-training, and memory as an architectural problem.

I prefer to start from first principles, go after the constraint that actually limits a model, and carry the idea through algorithms, training, and systems. I studied mathematics and physics at Weiyang College, Tsinghua University.

Interests

  • LLM Reasoning
  • Reinforcement Learning
  • Model Memory
  • Scientific LLMs

Education

  • Tsinghua University M.S. in Artificial Intelligence, Department of Computer Science and Technology · 2025.7 – present Knowledge Engineering Group (KEG). Advisor: Prof. Jie Tang.
  • Weiyang College, Tsinghua University B.S. in Basic Science of Mathematics and Physics · 2021.9 – 2025.7 GPA 3.85 / 4.00.

Experience

  1. 2023.7 – present

    Research Intern, ZhipuAI

    Research on LLM reasoning, post-training, and agents. Led work on process-rewarded tree search and later on self-training under low reward variance.

  2. 2023.6 – present

    Research, KEG, Tsinghua University

    Joined SciGLM / SciInstruct, then continued on reasoning, self-training, evaluation, and architecture-level memory.

  3. 2023.6 – 2023.7

    Oxford Summer Programme

    Coursework in AI/ML and a project on image classification with residual networks and attention.

Publications

† equal contribution. Leading papers are marked.

  1. Memory for Large Language Models

    Sining Zhoubian, Dan Zhang, Evgeny Kharlamov, Jie Tang

    arXiv 2026 · IEEE TPAMI (under review) First author

    An architecture-centric taxonomy of LLM memory along representation, update, and persistence — a map of how models store, route, and consolidate information.

    Paper

  2. Deeper is Not Always Better: Mitigating the Alignment Tax via Confident Layer Decoding

    Xuanming Zhang†, Sining Zhoubian†, Yuxuan Chen, Tianyi Tang, An Yang, Sean Du, Chujie Zheng, Fei Huang, Dayiheng Liu, Gao Huang, Jingren Zhou

    EMNLP 2026 Equal contribution

    Shows that later layers can perturb already-refined predictions (an alignment tax) and selects a nearer layer at an entropy valley, with no extra training.

    Paper Code

  3. ReST-RL: Achieving Accurate Code Reasoning of LLMs with Optimized Self-Training and Decoding

    Sining Zhoubian, Dan Zhang, Jie Tang

    arXiv 2025 · under review First author

    Addresses GRPO's vanishing reward variance with ReST-GRPO data selection, then guides search at test time with a value model (VM-MCTS).

    Paper Code

  4. DataSciBench: An LLM Agent Benchmark for Data Science

    Dan Zhang, Sining Zhoubian, Min Cai, Fengzu Li, Lekang Yang, Wei Wang, Tianjiao Dong, Ziniu Hu, Jie Tang, Yisong Yue

    ACL 2026 Findings Core contributor

    A benchmark for multi-step data-science agents, with a hierarchical Task–Function–Code evaluation that can be checked programmatically.

    Paper Code

  5. ReST-MCTS*: LLM Self-Training via Process Reward Guided Tree Search

    Dan Zhang†, Sining Zhoubian†, Ziniu Hu†, Yisong Yue, Yuxiao Dong, Jie Tang

    NeurIPS 2024 Equal contribution

    Infers process rewards from final correctness and uses MCTS* to collect traces for iterative self-training. 500+ citations.

    Paper Code Project

  6. SciInstruct: a Self-Reflective Instruction Annotated Dataset for Training Scientific Language Models

    Dan Zhang, Ziniu Hu, Sining Zhoubian, Zhengxiao Du, Kaiyu Yang, Zihan Wang, Yisong Yue, Yuxiao Dong, Jie Tang

    NeurIPS 2024 D&B Core contributor

    A critic-and-revise pipeline for college-level scientific instruction data, used to train the SciGLM models.

    Paper Code

  7. Rock Classification Based on Residual Networks

    Sining Zhoubian, Yuyang Wang, Zhihuan Jiang

    arXiv 2024 First author

    An undergraduate study of rock image classification with residual networks.

    Paper

Honors & Awards