CV

Contact Information

Name Kun Liang
Professional Title Ph.D. Student in Computer Science
Email kliang25@stu.pku.edu.cn
Location Beijing, China

Experience

  • 2025 - Present

    Beijing, China

    Research Intern
    Tencent Hunyuan
    RL Team; later Model Integration Team.
    • Researched RL and on-policy distillation (OPD) methods for post-training reasoning models.
    • Explored emerging post-training paradigms that aim beyond current practices, seeking more scalable and general recipes for improving LLM capabilities.

Education

  • 2025 - 2030

    Beijing, China

    Ph.D.
    Peking University
    Computer Science
    • Supervised by Prof. Yunfang Wu. Expected graduation June 2030.
    • Research interests: LLM post-training, with a focus on RL and broader on-policy learning for reasoning.
  • 2021 - 2025

    Beijing, China

    B.S.
    Beijing Normal University
    Computer Science
    • Major GPA: 3.82/4.0
    • First-Class Scholarship

Projects

  • ORBIT: Controllable Multi-Budget Reasoning Training Framework

    Tencent & Peking University. Advisors: Dr. Weijie Liu & Prof. Yunfang Wu.

    • Formulated reasoning allocation as a multi-budget reasoning problem and framed the training through an Exploration-Exploitation lens.
    • Exploration: designed a one-time context expansion-compression loop via multi-stage RL to discover Pareto-frontier reasoning behaviors at each budget level.
    • Exploitation: proposed multi-teacher on-policy distillation to fuse multiple frontier policies into a single unified model with well-separated, controllable reasoning modes.
    • Achieved controllable multi-budget reasoning and SOTA reasoning density, while preserving clear mode separation and high per-mode performance.
  • ADWIN: Adaptive Windows for Horizon-Aware On-Policy Distillation

    Tencent & Peking University. Advisors: Dr. Weijie Liu & Prof. Yunfang Wu.

    • Formulated OPD horizon selection as a prefix-admissibility problem, identifying when prefix-level teacher supervision is sufficiently reliable and faithful to preserve the full-rollout optimization direction, so that training can proceed synchronously.
    • Proposed ADWIN, a dual-path framework combining synchronous prefix updates with asynchronous full-rollout probes to audit prefix–full alignment and adapt the supervision horizon online.
    • Reduced end-to-end training cost by up to 4.1× while matching or outperforming full-rollout OPD across math and code reasoning benchmarks in both same-size and strong-to-weak settings.
  • Privy to the Foil: Recasting Value Estimation with a Self-Privileged Critic for RLVR

    Tencent & Peking University. Advisors: Dr. Weijie Liu & Prof. Yunfang Wu.

    • Recast value estimation in RLVR as an information problem, using verified reasoning evidence to help critics assess intermediate progress.
    • Proposed πPPO, which conditions the critic on leave-target-out, verified same-prompt rollouts as contrastive evidence while preserving a valid baseline and standard policy conditioning.
    • Improved value-estimation quality and mathematical reasoning performance over actor-critic and critic-free baselines, including with substantially smaller asymmetric critics.

Publications

Languages

English : TOEFL iBT 106 (R30 / L29 / S23 / W24)