Zisu Huang

Hi! I'm Zisu Huang (黄子骕), a first-year M.S. student in Computer Science at Fudan University, and a member of Fudan NLP Group, advised by Assoc. Prof. Xiaoqing Zheng and Prof. Xuanjing Huang. I earned my B.E. in Software Engineering from Fudan University in 2025. I am currently a research intern at Microsoft Research Asia working with Yifan Yang.

My research centers on AI agents — building capable, autonomous agents that can self-improve, and enabling them to reliably solve complex real-world tasks. In particular, I focus on:

  • Personalized agents: developing more effective personalized agents through both model training and external system optimization, such as persistent memory.
  • Agent evolution:
    • Parametric evolution: improving agentic capabilities through parameter updates, approached from both data-centric and algorithm-centric perspectives.
    • Non-parametric evolution: expanding the capability boundary of agents in a training-free manner, e.g., through agent skills, orchestration, and harness design.

News

  1. Two papers accepted to EMNLP 2026 — on query-specific rubric generation for DeepResearch agents and low-resource Sawndip–Chinese translation.

  2. Released two new works on agent skills — SkillOpt and SkillLens. SkillOpt sparked a lot of community discussion — thanks for the support!

  3. TRIP-Bench accepted to ICML 2026.

  4. New survey out — Reward Hacking in the Era of Large Models: Mechanisms, Emergent Misalignment, Challenges.

  5. SteeM accepted to ACL 2026 (Main Conference).

  6. Started a research internship at the Visual Computing Group, Microsoft Research Asia (Shanghai).

  7. RECAST accepted to ICLR 2026.

  8. Two papers accepted to EMNLP 2025 — SATER and FedQSN.

Selected Papers ( * denotes equal contribution )

  • SkillOpt teaser
    arXiv 2026 · Under Review

    Agent Skills Self-Evolution

    SkillOpt: Executive Strategy for Self-Evolving Agent Skills

    Yifan Yang*, Ziyang Gong*, Weiquan Huang*, Qihao Yang*, Ziwei Zhou*, Zisu Huang*, et al.

    10k+ GitHub stars 1M+ views 🤗 #1 Paper of the Day

    Paper Website Code Video

  • SkillLens overview
    arXiv 2026 · Under Review

    Agent Skills Empirical Study

    From Raw Experience to Skill Consumption: A Systematic Study of Model-Generated Agent Skills

    Zisu Huang*, Jingwen Xu*, Yifan Yang, Ziyang Gong, Qihao Yang, et al.

    Paper Website Code

  • Controllable Memory Usage
    ACL 2026 · Main

    Personalized Agents Long-Term Memory

    Controllable Memory Usage: Balancing Anchoring and Innovation in Long-Term Human–Agent Interaction

    Zisu Huang*, Muzhao Tian*, Xiaohua Wang, Jingwen Xu, Zhengkang Guo, et al.

    Paper Code

  • TRIP-Bench
    ICML 2026 · Poster

    Interactive Agents Deep Planning

    TRIP-Bench: A Benchmark for Long-Horizon Interactive Agents in Real-World Scenarios

    Yuanzhe Shen*, Zisu Huang*, Zhengyuan Wang*, Muzhao Tian*, et al.

    Paper

  • DeepResearch Bench performance with query-specific rubric rewards
    EMNLP 2026 · Main

    DeepResearch Agents Query-Specific Rubrics

    Learning Query-Specific Rubrics from Human Preferences for DeepResearch Report Generation

    Changze Lv, Jie Zhou, Wentao Zhao, Jingwen Xu, Shihan Dou, Zisu Huang, et al.

    Paper

  • RECAST
    ICLR 2026 · Poster

    Instruction Following Data Synthesis

    RECAST: Strengthening LLMs' Complex Instruction Following with Constraint-Verifiable Data

    Wenhao Liu, Zhengkang Guo, Mingchen Xie, Jingwen Xu, Zisu Huang, et al.

    Paper Code

  • Query Refinement RL
    arXiv 2024

    Query Refinement Reinforcement Learning

    Enhancing the Capability and Robustness of Large Language Models through Reinforcement Learning-Driven Query Refinement

    Zisu Huang*, Xiaohua Wang*, Feiran Zhang, Zhibo Xu, Cenyuan Zhang, et al.

    Paper Code

See the full list on my Google Scholar.

Projects ( * denotes equal contribution )

  • Reward Hacking survey
    Survey 2026

    Reward Hacking Reinforcement Learning

    Reward Hacking in the Era of Large Models: Mechanisms, Emergent Misalignment, Challenges

    Xiaohua Wang*, Muzhao Tian*, Yuqi Zeng*, Zisu Huang*, Jiakang Yuan*, Bowen Chen*, et al.

    Paper

views