Ph.D. Student  /  National University of Singapore

Leheng Sheng

I work on pretraining and mid-training of large language models, and study how they reason, how to keep them aligned, and what their internal representations can do.

Currently a research intern at Tencent Hunyuan (Code Pretrain).

Previously M.S. in Fintech at Tsinghua University and B.E. in Computer Science at Southeast University.

Email me Google Scholar GitHub CV
01

Research

i.

Pretraining & Midtraining

Building reasoning and coding abilities into large language models during pretraining and mid-training, including long-context and pretrain-stage reasoning.

ByteDance Seed · Tencent Hunyuan

ii.

Reasoning

How large reasoning models plan their thinking budget, learn from self-evolving rubrics, and reason over long contexts with gated memory.

NeurIPS ’25 · ICML ’26 · arXiv ’26

iii.

Alignment & Representations

Steering behaviour through a model’s internal representations, from refusal steering under null-space constraints to safety alignment with lightweight RL.

ICLR ’26 · ICLR ’26

02

News

  1. Joined Tencent Hunyuan (Code Pretrain) as a research intern through the Qingyun Talent Program.

  2. Self-Evolving Rubrics and When to Memorize released; When to Memorize accepted by ICML 2026.

  3. AlphaSteer accepted by ICLR 2026.

  4. On Reasoning Strength Planning accepted by NeurIPS 2025.

  5. Joined ByteDance Seed (Pretrain Team) as a research intern.

  6. Language Representations Can Be What Recommenders Need accepted by ICLR 2025 as an Oral.

03

Selected Papers

First and co-first author. * co-first, † corresponding.

  1. arXiv 2026

    Reinforcing Chain-of-Thought Reasoning with Self-Evolving Rubrics

    Leheng Sheng*, Wenchang Ma*, Ruixin Hong, Xiang Wang, An Zhang, Tat-Seng Chua

    arXiv

  2. ICML 2026

    When to Memorize and When to Stop: Gated Recurrent Memory for Long-Context Reasoning

    Leheng Sheng, Yongtao Zhang, Wenchang Ma, Yaorui Shi, Ting Huang, Xiang Wang, An Zhang, Ke Shen, Tat-Seng Chua

    arXiv

  3. ICLR 2026

    AlphaSteer: Learning Refusal Steering with Principled Null-Space Constraint

    Leheng Sheng*, Changshuo Shen*, Weixiang Zhao, Junfeng Fang, Xiaohao Liu, Zhenkai Liang, Xiang Wang, An Zhang, Tat-Seng Chua

    arXiv

  4. NeurIPS 2025

    On Reasoning Strength Planning in Large Reasoning Models

    Leheng Sheng, An Zhang, Zijian Wu, Weixiang Zhao, Changshuo Shen, Yi Zhang, Xiang Wang, Tat-Seng Chua

    arXiv

  5. ICLR 2025 Oral

    Language Representations Can Be What Recommenders Need: Findings and Potentials

    Leheng Sheng, An Zhang†, Yi Zhang, Yuxin Chen, Xiang Wang, Tat-Seng Chua

    arXiv Code

  6. SIGIR 2024

    On Generative Agents in Recommendation

    An Zhang*, Yuxin Chen*, Leheng Sheng*, Xiang Wang†, Tat-Seng Chua

    arXiv Code

  7. NeurIPS 2023

    Empowering Collaborative Filtering with Principled Adversarial Contrastive Loss

    An Zhang*, Leheng Sheng*, Zhibo Cai†, Xiang Wang, Tat-Seng Chua

    arXiv Code

  8. ICASSP 2023

    BrainNetFormer: Decoding Brain Cognitive States with Spatial-Temporal Cross Attention

    Leheng Sheng, Wehan Wang, Zhiyi Shi, Jichao Zhan, Youyong Kong†

    Paper

All 19 papers
04

Experience

Education

  1. National University of Singapore

    Ph.D. student, School of Computing

    2024 — now

  2. Tsinghua University

    M.S. in Finance (Fintech), School of Economics and Management

    2022 — 2024

  3. Southeast University

    B.E. in Computer Science, Chien-Shiung Wu College (Honors)

    2018 — 2022

Research & Industry

  1. Tencent Hunyuan

    Research Intern, Code Pretrain · Qingyun Talent Program

    Code pretraining for large language models.

    2026 — now

  2. ByteDance Seed

    Research Intern, Pretrain Team

    General & long-context LLM reasoning; pretrain-stage reasoning.

    2025 — 2026

  3. NExT++ Lab, NUS

    Research Intern

    LLM-based agent simulation for recommendation; a large-scale LLM agent simulation platform.

    2023 — 2024

  4. CITIC Securities

    Quantitative Research Assistant, Research Dept.

    Fund style classification.

    2021