Biography
I am Shenzhi Wang (็ๆ ๆง in Chinese), a Ph.D. candidate at LEAP Lab in the Department of Automation at Tsinghua University. My research focuses on post-training for foundation models, with an emphasis on reinforcement learning.
I am currently with Moonshot AI, where my recent work includes contributions to Kimi K3, focusing on scalable agentic post-training. Previously, I served as a research intern on Alibaba Qwen’s Post-training Team, where I worked on reinforcement learning, scalable post-training, and multimodal reasoning. I developed Beyond the 80/20 Rule and HopChain, with HopChain serving as one of Qwen3.5’s vision-language RLVR training tasks.
My research has received 
- Foundation Model Post-Training
- Reinforcement Learning
- LLMs, VLMs, and DLMs
Ph.D. in Artificial Intelligence, 2021โ2027 (expected)
Department of Automation, Tsinghua University
B.Eng. in Computer Science and Technology, 2017โ2021
SHENYUAN Honors College, Beihang University
News
- July 2026 โ The Flexibility Trap received the ๐ ICML 2026 Outstanding Paper Award.
- July 2026 โ Kimi K3 was released and open-sourced.
- March 2026 โ HopChain introduced scalable multi-hop data synthesis for generalizable vision-language reasoning.
- February 2026 โ Qwen3.5 was released and open-sourced.
- November 2025 โ Beyond the 80/20 Rule ranked among the top 5 most-cited NeurIPS 2025 papers.
- November 2025 โ Absolute Zero ranked among the top 25 most-cited NeurIPS 2025 papers.
- July 2025 โ OS Agents was selected as an ACL Oral paper.
- July 2025 โ PopAlign appeared at ACL.
- April 2025 โ Model Surgery appeared at NAACL.
- February 2025 โ Released EasyR1, an efficient and scalable multimodal reinforcement-learning framework.
- February 2025 โ Released the Xwen-Chat series, with Xwen-72B-Chat surpassing the then state-of-the-art DeepSeek-V3 in chat performance.
- February 2025 โ DiveR-CT was selected as an AAAI Oral paper.
- August 2024 โ ReCon appeared in Findings of ACL.
- August 2024 โ PsychoGAT appeared at ACL.
- May 2024 โ Released the Llama3-Chinese-Chat series, which later accumulated 1M+ downloads and reached Hugging Face Trending #7.
- December 2023 โ FamO2O was selected as a NeurIPS Spotlight paper.
- November 2023 โ GORL was published online in IEEE TNNLS.
- July 2023 โ OAP appeared at ICML.
- October 2021 โ Received the SenseTime Scholarship, awarded to 21 undergraduates nationwide.
- June 2021 โ Glancing at the Patch appeared at CVPR.
Featured Work
For a complete list of publications, please see my Google Scholar profile.
Kimi Technical Report ยท 2026
Qwen ยท 2026
One of the training tasks used in Qwen3.5 Multimodal RLVR ยท 2026Synthesizes instance-grounded multi-hop visual reasoning data for RLVR, improving generalization across 20 of 24 benchmarks on Qwen3.5-35B-A3B and Qwen3.5-397B-A17B.
Identifies arbitrary-order generation as a bottleneck for dLLM reasoning and introduces JustGRPO, an ICML 2026 Outstanding Paper Award-winning method that preserves parallel decoding while achieving state-of-the-art results.
NeurIPS 2025 ยท Top 5 Most-CitedShows that high-entropy minority tokens drive RLVR and develops selective policy-gradient updates that match or surpass full-gradient training; ranked among the five most-cited NeurIPS 2025 papers.
Hugging Face ยท 2025Post-trains Qwen2.5 into open 7B and 72B chat models, with Xwen-72B-Chat surpassing the then state-of-the-art DeepSeek-V3 in chat performance.
Hugging Face ยท 2024Delivers bilingual Chinese-English Llama 3 chat models with stronger role-play, tool-use, and math capabilities.
NeurIPS 2023 ยท SpotlightLearns a family of policies and state-adaptively selects improvement-constraint balances, consistently improving offline-to-online reinforcement learning across standard baselines.
Academic Service & Public Impact
- Reviewer for NeurIPS, ICML, ICLR, ACL, CoRL, ICRA, etc.
- Featured in a Chinese-language Z Potentials interview about research and academic growth.
- I share technical and research content with 2k+ followers on X, 2k+ on Xiaohongshu, and 10k+ on Zhihu.
Contact
For research discussions and collaboration, please feel free to reach out by email.
- wangshenzhi99@gmail.com