CV
Last updated: October 2026
Seeking: LLM post-training research internship (reinforcement learning / coding agents / general agents) in Beijing, available from early December 2026 for 3+ months.
Education
M.Eng. in Electronic Information (Computer Science and Technology), Tsinghua University, Expected June 2028
- Advisor: Associate Professor Jianfei Chen
- Research group: TSAIL, led by Professor Jun Zhu
B.Eng. in Computer Science and Technology, Tsinghua University, June 2026
- Major GPA: 3.8/4.0
- Relevant coursework (grade A or above): Ordinary Differential Equations; Probability and Statistics; Fundamentals of Programming; Programming and Training; Software Engineering; Operating Systems; Principles of Signal Processing; Database Special Topic Training
- Previous major: Medicine. Transitioned to Computer Science to pursue a passion for artificial intelligence.
Research Statement
- My research focuses on post-training of large models (reinforcement learning and data synthesis) and on agents, with the goal of improving capabilities on long-horizon tasks such as long-context reasoning, long-form generation, and complex coding. I also transfer effective reinforcement learning techniques from LLMs to diffusion and video generation models.
Publications
LongZero: Enhancing Long-Context Reasoning in LLMs from Zero Data
Zhennan Wan, Jianfei Chen
Under review. [arXiv] [Code]Scaling Long-Form Story Generation via Narrative State Tracking
Zhennan Wan, Jianfei Chen
Under review. [arXiv] [Code]Scaling External Knowledge Input Beyond Context Windows of LLMs via Multi-Agent Collaboration
Zijun Liu*, Zhennan Wan*, Peng Li, Ming Yan, Fei Huang, Yang Liu (*equal contribution)
ACL 2026 Main Conference. [Paper] [Code] [Data]
Research Experience
- September 2025 - Present: Research Member, Tsinghua Statistical Artificial Intelligence and Learning Group (TSAIL), Tsinghua University
- Reinforcement Learning for Diffusion Models (ongoing)
- Transferring techniques proven effective in LLM reinforcement learning to diffusion models.
- Systematically comparing step-wise likelihood methods (Flow-GRPO) with regression-based methods (DiffusionNFT, AWM) to design RL algorithms that are more efficient, more stable, and more robust to reward overfitting.
- LongZero: Enhancing Long-Context Reasoning in LLMs from Zero Data (first author, under review)
- Extended zero-data reinforcement learning from short-context to long-context tasks: the model constructs its own multi-hop reasoning and multi-round retrieval tasks, expands them into long-context data, and trains itself with GRPO.
- The method generalizes to general long-context reasoning, outperforming RL baselines trained on data synthesized from external corpora across 2 models and 5 benchmarks.
- Scaling Long-Form Story Generation via Narrative State Tracking (first author, under review)
- Studied length scaling in long-form fiction generation and designed an evaluation method for narrative consistency.
- Proposed NstAgent, an agent framework in which the model explicitly tracks structured narrative states, scaling story generation from 10,000 to 100,000 words without noticeable loss in consistency or writing quality.
- With reinforcement learning, a 4B model matched its 9B counterpart from the same model family.
- Reinforcement Learning for Diffusion Models (ongoing)
- September 2024 - May 2025: Research Intern, Natural Language Processing and Social Humanities Computing Lab (THUNLP), Tsinghua University
- Scaling External Knowledge Input Beyond Context Windows of LLMs via Multi-Agent Collaboration
- Developed ExtAgents, a multi-agent framework that distributes external knowledge across agents and enables intensive inter-agent communication, scaling external knowledge input beyond the context window without long-context training.
- Built the multi-hop QA benchmark ∞Bench+; on it and on tasks such as long survey generation, ExtAgents significantly outperformed existing training-free methods at the same input size, with performance improving as more knowledge is provided.
- Scaling External Knowledge Input Beyond Context Windows of LLMs via Multi-Agent Collaboration
- February 2024 - August 2024: Research Intern, 3D Visual Computing and Machine Intelligence (3DVICI) Lab, Institute for Interdisciplinary Information Sciences, Tsinghua University
- Editing Human Videos for Robotic Skill Training
- Used video editing techniques to convert human hand-object interaction videos into robotic hand-object interaction videos.
- Extracted 6D poses from human videos and compared human and robot data, referencing pipelines such as OpenVLA to evaluate the usefulness of human videos for robot training.
- Editing Human Videos for Robotic Skill Training
Internship Experience
- February 2026 - May 2026: LLM Algorithm Intern, TRAE Intelligent Coding Algorithms, ByteDance
- Seed Code Model - Coding Capability Optimization (MoE model with hundreds of billions of parameters, trained on hundreds of GPUs)
- Systematically evaluated long-horizon coding performance on public and internal benchmarks, including LoCoBench and LoCoDiff, to identify gaps versus leading models.
- Analyzed failure modes in code-state tracking and synthesized training data from real online TRAE trajectories, improving benchmark performance by approximately 8% to match GLM-5; integrated the pipeline into the model’s RL training framework to support data synthesis from dynamic rollouts.
- Built a systematic rubric-based evaluation framework for open-ended technical QA covering 10+ task types, 20+ domains, 6 general dimensions, and 4 task-specific dimensions; conducted targeted synthetic-data supervised fine-tuning on weak tasks, domains, and dimensions, narrowing the gap to GPT-5.2 by approximately 49% on a real-user QA dataset.
- Seed Code Model - Coding Capability Optimization (MoE model with hundreds of billions of parameters, trained on hundreds of GPUs)
- June 2025 - August 2025: Applied Research Intern, Code Intelligence Center, Technology and Engineering Group (TEG), Tencent
- CodeLLM Applications for Cursor Prediction and Smart Rewrite - Project Lead
- Synthesized data for 5 cursor-prediction and 15 smart-rewrite scenarios, producing 21K high-quality training examples.
- Conducted supervised fine-tuning and evaluation across Qwen2.5-Coder 0.5B/3B/7B/14B models; iterative improvements brought task performance on par with GPT-4.1.
- Built a VS Code extension with real-time inference and visual interaction; the project ranked 1st among 5 teams.
- CodeLLM Applications for Cursor Prediction and Smart Rewrite - Project Lead
Course Project
- April 2025 - June 2025: Frontiers in AI Safety and Governance (Spring 2025)
- Adaptive Safety Priming: Inference-Time Safeguards for Large Reasoning Models
- Developed Adaptive Safety Priming (ASP), a lightweight, dynamic safety mechanism that leverages the step-by-step inference process of large reasoning models to enable real-time intervention. This approach provides a more adaptive and resource-efficient path toward robustly safe models. [Report]
- Adaptive Safety Priming: Inference-Time Safeguards for Large Reasoning Models
Skills
- Programming: C, C++, Python
- Machine learning: PyTorch, vLLM, VeRL, FSDP
- Tools: Git, Linux, Docker, LaTeX
- English: College English Test Band 6 (CET-6): 601/710
Awards
- Academic Excellence Scholarship, 2024-2025 Academic Year
- Tsinghua University Software Engineering Outstanding Project Award, 2024