Junjie Wang (王军杰)

Junjie Wang (王军杰)

Postdoctoral Fellow

About me

Junjie Wang avatar

Junjie Wang (王军杰) is currently a postdoctoral researcher at the IIGROUP Lab, Tsinghua University, supervised by Prof. Yujiu Yang. His research focuses on Learning to Reason for Self-Evolving Agents, spanning knowledge-grounded reasoning in foundation models, interaction-based reasoning in digital agents, and physical-world reasoning in embodied agents. He received his Ph.D. in Engineering from Waseda University, under the supervision of Prof. Tetsuya Sakai.

Email: wangjunjie@sz.tsinghua.edu.cn (Please state your purpose / 请您注明来意)

📰 News / 最近

  • [09/2026] 📜 3 papers are accepted to NeurIPS 2026!
  • [08/2026] 📜 8 papers are accepted to EMNLP 2026 !
  • [06/2026] 🏅 Selected as an ICML 2026 Golden Reviewer.
  • [04/2026] 📜 1 paper is accepted to ICML 2026 !
  • [04/2026] 📜 2 papers are accepted to ACL 2026 !
  • [03/2026] 📜 1 paper is accepted to TMLR ! 获选CAAI-腾讯犀牛鸟研究计划 !
  • [01/2026] 🎉 ROCO@AAAI2026 决赛冠军: URL
  • [01/2026] 📜 1 paper is accepted to ICLR 2026 !
  • [11/2025] 📜 1 paper (Oral) is accepted to AAAI 2026 !

🌌 Research Vision / 研究愿景

Learning to Reason: From Knowledge to Interaction to Physical Intelligence

学会推理：从知识驱动，到交互驱动，再到物理世界智能

My central question is: How can reasoning become a continually improving capability? I study a progression from knowledge-grounded reasoning to interaction-based reasoning and physical-world reasoning, with decreasing reliance on human-defined knowledge and increasing agent-driven information acquisition.

我的核心问题是：如何让推理成为一种能够持续改进的能力？ 我关注从知识驱动推理、交互驱动推理到物理世界推理的演进：对人类预定义知识的依赖逐步减少，智能体通过交互主动获取信息的能力逐步增强。

Three stages from foundation models with knowledge-grounded reasoning, through digital agents with interaction-based reasoning, to embodied agents with physical-world reasoning; human-defined knowledge decreases as agent-driven information acquisition increases

From foundation models to digital agents and embodied agents: human knowledge → interactive feedback → physical consequences.
从基础模型，到数字智能体，再到具身智能体：人类知识 → 交互反馈 → 物理世界后果。

🧭 Three Stages of Reasoning / 推理的三个阶段

  1. Knowledge-Grounded Reasoning / 知识驱动推理 — Develop reasoning capabilities in foundation models through pretraining on human-curated data and knowledge.
    通过在人类整理的数据与知识上进行预训练，发展基础模型的推理能力。

  2. Interaction-Based Reasoning / 交互驱动推理 — Enable digital agents to actively acquire information through environments and tools, verify feedback, and turn interaction experience into capability improvements.
    让数字智能体通过环境与工具主动获取信息、验证反馈，并将交互经验转化为能力提升。

  3. Physical-World Reasoning / 物理世界推理 — Enable embodied agents to connect perception, world modeling, prediction, and action, using planning and physical consequences to inform subsequent decisions.
    让具身智能体将感知、世界建模、预测与行动形成闭环，通过规划指导行动，并根据物理世界中的行动后果调整决策。

🔄 Toward Reasoning-Centric RSI / 迈向以推理为中心的 RSI

Interaction → Verification → Attribution → Internalization → Capability Improvement ↺

My long-term goal is Reasoning-Centric Recursive Self-Improvement (RSI): agents that can verify which experiences are trustworthy, attribute success and failure, internalize high-value lessons, and eventually improve the mechanisms by which they improve.

我的长期目标是构建以推理为中心的递归自我改进（Reasoning-Centric RSI）：让智能体能够验证经验、归因成败、内化高价值反馈，并最终改进其自身的改进机制。

✒️ Recent Professional Services / 最近的专业服务

  • Golden Reviewer: ICML 2026
  • Area Chair: ACL ARR 2025 October (EACL); ACL ARR 2026 January (ACL); ACL ARR 2026 March (EMNLP).
  • Senior PC member: WSDM 2026
  • Workshop Organiser: SIGIR-AP 2025 BREV-RAG Workshop

Publications

Details in Publications Page

Total Publications: 40

  • ⭐: Co-first Author
  • 🚩: Corresponding Author
  • 💭: Under Review