Seek Before You Move: Evidence Seeking for Progress Grounding in Vision-Language Navigation
arXiv preprint, 2026
TL;DR: Learns when to gather visual evidence before acting to improve progress grounding in long-horizon navigation.
Zhimin Wang 王智民
Ph.D. student · Electronic Information · Tsinghua University
I am advised by Prof. Zhi Wang and work with MMLab@SIGS in the Agentic and Embodied Large Model Training Group. I received my bachelor’s degree from Shandong University in 2025.
My research focuses on long-horizon decision making for language model agents, including vision-language navigation, multi-agent collaboration, and GUI agents.
News
- SeekVLN is available on arXiv. Project page.
- DeCoPatch was accepted to ECCV 2026.
- Learning with Challenges is available on arXiv.
- Generalist Policies for Network Optimization is available on arXiv.
- CoBel-World is available on arXiv.
Publications
* Equal contribution
DeCoPatch: Revealing Causal Latent Subspaces in Vision-Language Models for GUI Grounding
ECCV 2026 · Accepted [THU-A]
TL;DR: Identifies low-dimensional causal mechanisms in VLMs and applies sparse activation interventions for GUI grounding.
Learning with Challenges: Adaptive Difficulty-Aware Data Generation for Mobile GUI Agent Training
arXiv preprint, 2026 · Under review [CCF-A]
TL;DR: Profiles agent capabilities and generates tasks and trajectories with adaptive structural and semantic difficulty.
Large Language Models as Generalist Policies for Network Optimization
arXiv preprint, 2025 · In submission to Nature Communications
TL;DR: Aligns LLMs to networking tasks and routes simpler control cases to a lightweight policy.
CoBel-World: Harnessing LLM Reasoning to Build a Collaborative Belief World for Optimizing Embodied Multi-Agent Collaboration
arXiv preprint, 2025 · Under review [CCF-A]
TL;DR: Uses LLM reasoning to model shared beliefs for embodied multi-agent collaboration with less coordination overhead.
Robot Navigation Demo
ELF-system · Oct. 2026
Language-guided, multi-stop indoor navigation with a quadruped robot.
Edited indoor footage with illustrative planning overlays and UI inserts. Some segments are accelerated; this is a demonstration, not a continuous system benchmark.
Click the image to play the demo.
Education
Sep. 2025–present
Tsinghua University
Ph.D. student in Electronic Information
Advisor: Prof. Zhi Wang · MMLab@SIGS, Agentic and Embodied Large Model Training Group
Sep. 2021–Jun. 2025
Shandong University
Bachelor’s degree
Experience
Jul.–Aug. 2026
Seres Group · Research Intern
Worked on a robot vision-language navigation system and automated navigation data generation using a high-quality 3D Gaussian dataset.
Honors & Awards
- National Scholarship for Undergraduates, 2021, 2022, 2023 — the only recipient in my major for three consecutive years.
- First Prize, China Undergraduate Mathematical Contest in Modeling — top 0.5% nationwide.
- Outstanding Student, Shandong University — top 3% university-wide.
- Outstanding Graduate, Shandong University — top 5% university-wide.
Academic Work
Teaching assistant for 64100033-200, Fundamentals of Big Data Systems (B), 2025.
Contact
wangzhim25@mails.tsinghua.edu.cn · 2391604943@qq.com
Information Building, Tsinghua Campus, University Town of Shenzhen, P. R. China