Shenzhi Wang王慎执
Post-training for agents and reasoning,
especially reinforcement learning.
I am a Ph.D. candidate at LEAP Lab, Tsinghua University, currently with Moonshot AI. Previously, I was a research intern on Alibaba Qwen’s Post-training Team.
My research spans foundation-model post-training and reinforcement learning, with 
Alongside this research, my recent team projects include Kimi K3.1 Pro and Kimi K3. In HopChain, I explore scalable data synthesis for vision-language reinforcement learning; it is one of the training tasks used in Qwen3.5.
I also build open models: Llama3-Chinese-Chat has 1M+ downloads and reached Hugging Face Trending #7; Xwen-Chat includes a 72B model that surpassed the then state-of-the-art DeepSeek-V3 in chat performance.
Selected Research
Research papers
Beyond the 80/20 Rule: High-Entropy Minority Tokens Drive Effective Reinforcement Learning for LLM Reasoning
Identifies high-entropy tokens as critical decision points in reasoning, accounting for nearly all performance gains from reinforcement learning with verifiable rewards.
The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models
Reveals the limits of arbitrary-order reasoning in diffusion language models and introduces JustGRPO, preserving parallel decoding while improving reasoning.
HopChain: Multi-Hop Data Synthesis for Generalizable Vision-Language Reasoning
Synthesizes multi-hop visual reasoning data for RLVR, improving generalization on 20 of 24 benchmarks across two Qwen3.5 models.
Team projects
Kimi K3.1 Pro
Scales post-training and joint reinforcement learning across domains to strengthen coding, reasoning, and long-horizon agents, with a one-million-token context window.
Kimi K3: Open Frontier Intelligence
A frontier model project advancing scalable agentic post-training and open intelligence.
Qwen3.5: Towards Native Multimodal Agents
A native multimodal model family; HopChain is one of its vision-language RLVR training tasks.
Open models & earlier work
Xwen-Chat Series
Open 7B and 72B chat models post-trained from Qwen2.5, with a focus on stronger conversational capabilities.
Llama3-Chinese-Chat Series
Bilingual Llama 3 chat models for Chinese and English, with improved role-play, tool use, and math capabilities. Over one million downloads.
Train Once, Get a Family: State-Adaptive Balances for Offline-to-Online Reinforcement Learning
Learns a family of policies and selects state-adaptive balances between policy improvement and constraints for offline-to-online reinforcement learning.
News
- October 2026Kimi K3.1 Pro was introduced, advancing coding and long-horizon agent capabilities through scaled post-training.
- July 2026Kimi K3 was released and open-sourced.
- July 2026The Flexibility Trap received the 🏆 ICML 2026 Outstanding Paper Award.
- March 2026HopChain introduced scalable multi-hop data synthesis for generalizable vision-language reasoning.
- February 2026Qwen3.5 was released and open-sourced.
- November 2025Beyond the 80/20 Rule ranked among the top 5 most-cited NeurIPS 2025 papers.
- November 2025Absolute Zero ranked among the top 25 most-cited NeurIPS 2025 papers.
- July 2025OS Agents was selected as an ACL Oral paper.
- July 2025PopAlign appeared at ACL.
- April 2025Model Surgery appeared at NAACL.
- February 2025Released EasyR1, an efficient and scalable multimodal reinforcement-learning framework.
- February 2025Released the Xwen-Chat series, with Xwen-72B-Chat surpassing the then state-of-the-art DeepSeek-V3 in chat performance.
- February 2025DiveR-CT was selected as an AAAI Oral paper.
- August 2024ReCon appeared in Findings of ACL.
- August 2024PsychoGAT appeared at ACL.
- May 2024Released the Llama3-Chinese-Chat series, which later accumulated 1M+ downloads and reached Hugging Face Trending #7.
- December 2023FamO2O was selected as a NeurIPS Spotlight paper.
- November 2023GORL was published online in IEEE TNNLS.
- July 2023OAP appeared at ICML.
- October 2021Received the SenseTime Scholarship, awarded to 21 undergraduates nationwide.
- June 2021Glancing at the Patch appeared at CVPR.
Education
2021–2027 · expected
Ph.D. in Artificial Intelligence
Department of Automation
Tsinghua University
2017–2021
B.Eng. in Computer Science and Technology
SHENYUAN Honors College
Beihang University
Service & Outreach
Academic Service
Reviewer for NeurIPS, ICML, ICLR, ACL, CoRL, ICRA, and related venues.
Writing & Outreach
I share research updates and technical notes on X, Zhihu, and Xiaohongshu.
Z Potentials interview · in Chinese
Contact
I’m always happy to connect, exchange ideas, and explore collaborations. Feel free to say hello!
wangshenzhi99 [at] gmail [dot] com