😎 Hi, I am Zhenyu Li (李振宇). I received my Ph.D. in Computer Science and Technology from Tsinghua University in 2026, advised by Prof. Jianyong Wang. Prior to that, I received my B.S. degree from Tsinghua University in 2021.
My primary research focuses on enhancing the reasoning and agentic capabilities of large language models:
- Native Agentic Model: Equip LLMs with intrinsic agentic capabilities such as instruction following, multi-step planning, and complex tool orchestration, towards models that can autonomously handle diverse real-world tasks.
- Agent RL: Extend reinforcement learning to open-ended, hard-to-verify agent scenarios by designing effective reward mechanisms (e.g., rubric-based evaluation), improving agent performance in deep research and creative design.
- Long-Context & Efficient Architecture: Extend LLMs’ long-context capabilities while designing efficient attention mechanisms and in-context updatable architectures for test-time learning.
📖 Educations
- [2021.09-2026.06] Ph.D. Tsinghua University (Advisor: Prof. Jianyong Wang; Outstanding Graduate)
- [2017.09-2021.07] B.S. Tsinghua University (Outstanding Graduate of the Department)
🔥 News
- 🔥 [2026.06.23]: We released Seed 2.1, a next-generation agent model built for real-world productivity, with significantly improved general agent and end-to-end coding capabilities. As a core contributor, I helped build the end-to-end pipeline for improving Seed’s capabilities under general harness — covering high-quality data reflow, query collection & filtering, Sandbox-based large-scale rollout on Hermes / OpenClaw frameworks, black-box Harness RL, and rapid capability evaluation — which now powers Harness optimization for nearly all mainstream frameworks on Volcano Engine. On our internal SeedClawBench, Seed2.1-Pro reaches 66.6, on par with GPT-5.5 and Claude-4.7-Opus, and ranks above Claude Opus 4.6 in large-scale anonymous developer preference tests under the Hermes / OpenClaw frameworks. More details check the model card.
- 🎉 [2026.06]: Efficient Attention Mechanisms for Large Language Models: A Survey has been accepted by Cell Patterns.
- 🔥 [2026.04.21]: Agent-World featured as 🤗 HF Daily Paper #2 ! It is the first self-evolving training arena for general agents. Agent-World-14B achieves 55.8% on BFCL-V4 (surpassing DeepSeek-V3.2-685B), 65.4% on tau²-Bench and outperforms all open-source baselines across 23 agent benchmarks. It exhibits strong generalization in General Reasoning, Agentic Search & Coding domains, and delivers significant improvements in advanced agentic assistant scenarios such as SkillsBench and ClawEval. More details refer to our paper and demo.
- 🎉 [2026.02.16]: Seed 2.0 made its debut in the LM Arena rankings, securing 6th place in the Text Arena and 3rd place in the Vision Arena respectively ! 🎉🎉🎉
- 🔥 [2026.02.14]: We released Seed 2.0. It is a powerful unified model with comprehensive capabilities surpassing Gemini 3 Pro with extreme superior agent capabilities. As a contributor, I am responsible for deep research ability (DeepConsult 61.1, DeepResearchBench 53.3, ResearchRubrics 50.7), which has been greatly improved compared to version 1.8. More details check the model card.
- 🚀 [2026.02]: Doubao Super Mode (豆包超能模式) is now fully available to all users! It is a powerful AI agent that autonomously handles complex tasks through multi-tool collaboration — integrating web browsing, multi-round search, code execution, data analysis, file processing, and multimodal content generation — supporting a wide range of scenarios including deep research, report writing, coding assistance, travel planning, and creative design. As a key contributor, I am primarily responsible for the reinforcement learning (RL) pipeline that powers its agentic capabilities.
- 🌟 [2025.12.18]: Our team is releasing Seed1.8, a model designed for generalized real-world agency. It supports text and image inputs and with its powerful multimodal processing capabilities, it demonstrates good performance across various complex application scenarios such as information retrieval, coding, Graphical User Interface (GUI) interaction. More details check the model card.
- 🎉 [2025.05]: FocusLLM has been accepted by ACL 2025 Main. A framework for extending LLM context length through dynamic condensing.
- 🎉 [2025.05]: Negative Matters has been accepted by ACL 2025 Main.
- 🎉 [2025.05]: Maximum Score Routing has been accepted by ACL 2025 Findings.
- 🎉 [2024.07]: UniMem has been accepted by COLM 2024.
- 🎉 [2023.12]: FlexKBQA has been accepted by AAAI 2024 (Oral, Top 2%).
- 🎉 [2023.11.23] We released the XAgent v1.0.0! 🎉🎉🎉
💌 Projects
- Awesome-Harness-Self-Improvement: A curated reading list on harness engineering for recursive self-improvement of LLM agents (EN/ZH).
- XAgent: An Autonomous Agent for Complex Task Solving.
📝 Selected Publications
- Seed 2.1: A Next-Generation Agent for Real-World Productivity
Seed Team
Model card 2026, [Paper] - Seed 2.0 Model Card: Towards Intelligence Frontier for Real-World Complexity
Seed Team
Model card 2026, [Paper] - Seed 1.8 Model Card: Towards Generalized Real-World Agency
Seed Team
Model card 2025, [Paper] - Efficient Attention Mechanisms for Large Language Models: A Survey
Yutao Sun*, Zhenyu Li*, Yike Zhang*, Tengyu Pan*, Bowen Dong*, Yuyi Guo, Jianyong Wang
Cell Patterns, [Paper] - FocusLLM: Precise Understanding of Long Context by Dynamic Condensing
Zhenyu Li, Yike Zhang, Tengyu Pan, Yutao Sun, Zhichao Duan, Junjie Fang, Rong Han, Zixuan Wang, Jianyong Wang
ACL 2025 Main, [Paper] - Negative Matters: Multi-Granularity Hard-Negative Synthesis and Anchor-Token-Aware Pooling for Enhanced Text Embeddings
Tengyu Pan, Zhichao Duan, Zhenyu Li, Bowen Dong, Ning Liu, Xiuxing Li, Jianyong Wang
ACL 2025 Main, [Paper] - Maximum Score Routing For Mixture-of-Experts
Bowen Dong, Yilong Fan, Yutao Sun, Zhenyu Li, Tengyu Pan, Xun Zhou, Jianyong Wang
ACL 2025 Findings, [Paper] - UniMem: Towards a Unified View of Long-Context Large Language Models
Junjie Fang, Likai Tang, Hongzhe Bi, Yujia Qin, Si Sun, Zhenyu Li, Haolun Li, Yongjian Li, Xin Cong, Yankai Lin, Yukun Yan, Xiaodong Shi, Sen Song, Zhiyuan Liu, Maosong Sun
COLM 2024, [Paper] - FlexKBQA: A Flexible LLM-Powered Framework for Few-Shot Knowledge Base Question Answering
Zhenyu Li, Sunqi Fan, Yu Gu, Xiuxing Li, Zhichao Duan, Bowen Dong, Ning Liu, Jianyong Wang
AAAI 2024 (Oral, Top 2%), [Paper] - Optimization Techniques for Unsupervised Complex Table Reasoning via Self-Training Framework
Zhenyu Li, Xiuxing Li, Sunqi Fan, Jianyong Wang
IEEE Transactions on Knowledge and Data Engineering (TKDE), [Paper] - Toward a Unified Framework for Unsupervised Complex Tabular Reasoning
Zhenyu Li, Xiuxing Li, Zhichao Duan, Bowen Dong, Ning Liu, Jianyong Wang
ICDE 2023, [Paper] - Effective Few-Shot Named Entity Linking by Meta-Learning
Xiuxing Li*, Zhenyu Li*, Zhengyan Zhang, Ning Liu, Haitao Yuan, Wei Zhang, Zhiyuan Liu, Jianyong Wang
ICDE 2022, [Paper]
Full paper list refer to google scholar.
🎖 Honors and Awards
- [2026] Outstanding Graduate, Tsinghua University
- [2024] Comprehensive Excellence Scholarship, Tsinghua University
- [2023] Comprehensive Excellence Scholarship, Tsinghua University
- [2022] Comprehensive Excellence Scholarship, Tsinghua University
- [2021] Outstanding Graduate of the Department of Computer Science, Tsinghua University
- [2021] National Encouragement Scholarship
- [2020] First Prize of Hua Luogeng Mathematical Contest in Modeling, Tsinghua University (Top 3%)
- [2019] Academic & Social Work Excellence Scholarship, Tsinghua University
- [2018] Second Prize of the Physics Competition for College Students in Parts of China
💻 Work Experience
- [2026.7-present] Bytedance-Seed, Researcher (Agent Group)
- [2025.9-2026.6] Bytedance-Seed Intern (Agent Group, supervised by Dr. Yujia Qin)
- [2025.6-2025.9] Bytedance-Seed Intern (Pretrain Group, supervised by Ke Shen)
- [2024.8-2025.4] Microsoft Research Asia, NLC Group Intern (Mentor: Dr. Li Dong)
- [2023.9-2024.2] Tsinghua University, THUNLP Group Research Collaboration (Advisor: Prof. Zhiyuan Liu)
🎓 Teaching Experience
- [2021.09-Present] Undergraduate Counselor, Tsinghua University (12·9 Counselor Scholarship)
- [2022.01-2023.12] Teaching Assistant of Software Engineering, Tsinghua University
📚 Academic Services
- Serving as a reviewer of ACL 2025, ICLR 2026
💡 Lifestyle
- My hobbies include but are not limited to 🎤 singing, 🏸 badminton and 🥏 frisbee.