About me
Junjie Wang (王军杰) is currently a postdoctoral researcher at the IIGROUP Lab, Tsinghua University, supervised by Prof. Yujiu Yang. His research interests include natural language processing, multimodal reasoning, and embodied intelligence. 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
- [01/2026] 🎉 ROCO@AAAI2026 决赛冠军: URL
- [01/2026] 📜 1 paper are accepted to ICLR 2026 !
- [11/2025] 📜 1 paper (Oral) are accepted to AAAI 2026 !
🌌 Long-term Goal / 长期目标
-> Practical Epistemology / 实践认识论 🧠
My long-term goal is to engineer (self-evolving) AGI as a form of Practical Epistemology: turning “knowing” into an operational, testable process. Concretely, I formalize cognition into structured representations, shape executable behaviors via post-training signal engineering, and make reasoning–action–creation verifiable through evidence, traces, and executable artifacts. “Self-evolving” means a reproducible closed loop—verification signals → training signals and data/task redesign → regression-tested capability gains—rather than unconstrained self-modification.
我的长期目标是将（自进化的）通用智能工程化为一种“实践认识论”：把“知道/理解”变成可操作、可检验、可复现的过程。具体而言，我将认知形式化为结构化表示，通过后训练的信号工程塑形可执行行为，并以证据链、过程轨迹与可执行产物对推理—行动—创造进行过程级验收。“自进化”指可复现闭环：验收信号 → 训练信号与数据/任务重构 → 回归测试下的能力累积，而不是无约束自我改写。
🔎 Research Interests / 研究兴趣
- Structured Cognition / 结构化认知
- Post-Training Behavior Shaping / 后训练行为塑形
- Process Verification & Diagnosis / 过程验收与诊断
- Self-Evolution Flywheel / 自进化飞轮
🌱 Derived Verticals / 落地方向 (AI for X)
- AI for Creation（创造/编辑）
- AI for Knowledge Work（知识/科学知识工作流）
- AI for Tool-Using Automation（工具/GUI/代码自动化）
- AI for Robots（机器人）
✒️ Recent Professional Services / 最近的专业服务
- Area Chair: ACL ARR 2025 October (EACL 2026)
- Senior PC member: WSDM 2026
- Workshop Organiser: SIGIR-AP 2025 BREV-RAG Workshop
Publications
Details in Publications Page
Total Publications: 28
- ⭐: Co-first Author
- 🚩: Corresponding Author
- 💭: Under Review
Research Interests / 研究兴趣 展开说明
- Structured Cognition / 结构化认知
- Task/interface unification: compositional formulations for reasoning, extraction, and selection
- 任务/接口统一：将推理、抽取与选择统一为可组合表述
- Multimodal interaction representations: structured grounding and cross-modal interaction modeling
- 多模态交互表征：结构化 grounding 与跨模态交互建模
- Reasoning primitives: multi-paradigm and cooperative reasoning as reusable cognitive mechanisms
- 推理原语：多范式与协作推理的可复用认知机制
- Post-Training Behavior Shaping / 后训练行为塑形
- Supervision signal engineering: mining, selection, and reweighting of high-value learning signals
- 监督信号工程：高价值学习信号的挖掘、筛选与重加权
- Tool/retrieval alignment: learning transferable tool-using and decision behaviors under data constraints
- 工具/检索对齐：在数据效率约束下学习可迁移的工具使用与决策行为
- Capability injection & transfer: behaviors that generalize across substrates (text/code/GUI/tools)
- 能力注入与迁移：跨载体（文本/代码/GUI/工具）的可泛化稳定行为
- Process Verification & Diagnosis / 过程验收与诊断
- Process-centric evaluation: stage-wise acceptance criteria beyond outcome-only scores
- 过程导向评测：以分阶段验收标准替代单一结果分数
- Executable verification interfaces: chart-to-code, long-horizon GUI traces, repository-level code tasks
- 可执行验证接口：chart-to-code、GUI 长链轨迹、仓库级代码任务
- Diagnostic taxonomies: structured failure modes for localization and regression testing
- 诊断分类体系：结构化失败模式以支持定位与回归测试
- Self-Evolution Flywheel / 自进化飞轮
- Verification → training transformation: converting diagnostics into objectives, curricula, and data/task redesign
- 验收信号→训练信号：将诊断转化为训练目标、课程与数据/任务重构
- Loop stability & milestones: measurable stages and regression suites for cumulative generalization
- 闭环稳定性与里程碑：用阶段性指标与回归测试集保障能力累积
- Infrastructure for evolution: datasets/benchmarks/systems enabling reproducible iteration
- 演化基础设施：数据/基准/系统资产支撑可复现迭代