Education

• Ph.D. in Artificial Intelligence, Zhejiang University, Sep 2022 – present
  • Member of REAL Lab, advised by Dr. Yongliang Shen
  • Coursework: Natural Language Processing, Knowledge Graphs, Knowledge Reasoning & Representation
• B.S. in Information Management and Information Systems, Beijing Normal University, Sep 2018 – Jun 2022
  • Minor: Data Science and Big Data Technology, Beijing Normal University, 2020 – 2022

Research Interests

• LLM post-training: SFT and RL recipes for language and multimodal models that improve reasoning, agentic, and subjective-alignment capabilities
• Reward modeling & reinforcement learning: reliable, scalable reward signals and verification mechanisms that support large-scale RL training
• AI agents: agents that make decisions and act reliably on complex, real-world tasks

Honors & Awards

• National Scholarship (国家奖学金)
• Beijing Merit Student (北京市三好学生)
• Beijing Outstanding Graduate (北京市优秀毕业生)

Internships

• Tencent, Hunyuan Foundation Model Team (Qingyun Program), Apr 2026 – present
  • VLM visual post-training
    • Contributed to merged SFT releases for VLM post-training and set up how capability tracks collaborate on them; the resulting model is the starting point for RL
    • Pairwise RL for subjective capabilities: about 4× faster RL training and a 10% gain on subjective tasks; merged into the production recipe
    • Rubric RL for subjective capabilities: explored how rubrics are produced and how judgments are optimized, further improving subjective tasks
    • Built VisualAgent capabilities for multimodal tasks through harness design, RL training, and self-distillation
    • Built the team toolbox for VisualAgent-based data synthesis, data quality checks, offline evaluation, and harness and tool iteration
• Ant Group, Ling Foundation Model Team (research intern), Aug 2025 – Apr 2026
  • Ring reasoning models
    • Built reasoning-effort control through joint SFT–RL optimization, supporting Low / Medium / High / xHigh reasoning depths
    • Improved agentic capability on software-engineering tasks with two-stage SFT–RL, gaining about 6% on SWE-Bench
    • Explored context compaction for agents: in long-CoT mode, 40% lower inference latency and 9% better results
    • Co-authored the Ring-1T technical report, covering the full training pipeline of a trillion-parameter reasoning model
• Meituan, LongCat Foundation Model Team, Jun 2023 – Jul 2025
  • Math reasoning
    • Owned math-reasoning data merges across pre-training, annealing, and SFT, delivering math data to the mainline model at each stage
    • Synthesized math instruction data for annealing, with diverse query synthesis and controllable response synthesis; 11% overall improvement
    • Pre-training, fine-tuning, and RL experience with Dense and MoE models at 1B / 7B / 100B+ scale
    • Built the team’s data-synthesis and evaluation framework for fast benchmark onboarding and quick idea validation
  • General SFT
    • Produced, cleaned, and versioned mainline instruction-tuning data, and designed schemes for merging specialized capabilities
    • Long context: integrated long-context evaluation and explored positional-encoding interpolation, extending the mainline context window about 8× with little loss
    • Tool use: built tool-calling support from scratch while preserving general capabilities

Skills

• LLM training: end-to-end pre-training and fine-tuning with Megatron-Core; Dense and MoE models from 1B to 100B+, including parallelism strategy, training stability, and throughput optimization
• Reinforcement learning: large-scale RL pipelines with veRL, from rollout and reward computation to policy updates; fast rollout and deployment with vLLM and SGLang
• Engineering: Python (multiprocessing / multithreading, async, decorators), Git collaboration and branching, Linux and shell scripting, multi-node multi-GPU jobs on clusters and debugging distributed programs
• Frameworks: designed and built evaluation frameworks (fast benchmark onboarding, distributed inference, automatic result aggregation) and data-synthesis frameworks (diverse query synthesis, controllable response generation, quality filtering)