Research interests
My current interests center on recursive self-improvement (RSI), with an open focus across agents and models. My previous research spans multimodal models, embodied intelligence, vision-language-action systems, and navigation.
Education
Peking University
B.S. in Computer Science and Technology
Research & industry experience
Research
Dec 2025 — Sep 2026
Center on Frontiers of Computing Studies, Peking University
Undergraduate Research Assistant
Advisor: Hao Dong
Research on post-training for vision-language-action systems and language-guided navigation.
Mar 2026 — Jun 2026
Wangxuan Institute of Computer Technology, Peking University
Research Rotation
Advisor: Yang Liu
Research on spatio-temporal video grounding through multi-agent debate.
Industry
Jun 2026 — Sep 2026
JD Technology
Embodied AI Algorithm Engineering Intern
Worked on long-horizon task data, chain-of-thought synthesis, distributed SFT/RL post-training with Qwen-VL models, evaluation, and failure analysis.
Nov 2025 — Mar 2026
Ubiquant Innovation
AI Agent Engineering Intern
Contributed to Framía and One, building capabilities for content understanding, page generation, visual layout, research, and complex office tasks.
Papers
ASCRIBE: Retrospective Responsibility Routing Post-Training for Hierarchical VLA System
One Scan is Enough: Demonstration-Free Adaptation for Language-Guided Navigation
DualFuture: Event-Triggered Dual-Future Residual Correction for Vision-Language-Action Policy
VERDICT: Unleashing Cross-Method Complementarity for Spatio-Temporal Video Grounding through Debate
* Equal contribution. † Corresponding author.
Selected projects
CritPT-RL
July 2026
An end-to-end GRPO post-training and evaluation pipeline for scientific coding tasks, with synthetic data, reward functions, and executable verifiers.
README-AI
February–May 2026
An AI paper reader for translating, discussing, and annotating research papers, with contextual figure explanations and related-work search.
Community
PLib · Core Organizer
Aug 2025 — Present
Help organize P-Talk events and develop P-Lib, a course notes and resources platform with more than 5,000 registered users.
Technical experience
- Training & evaluation
- Python, PyTorch, VLM/VLA supervised fine-tuning, GRPO/RL, reward design, model evaluation
- Data & applications
- Multimodal data cleaning, data synthesis, retrieval-augmented generation, agent workflows