publications

2026

1. arXiv

   

   UI-Mate: Advancing Open-Weight Foundation GUI Agents with In-Context Demonstrations

   Tencent Hy Team

   In arXiv preprint arXiv:2608.15930. Core contributor to the technical report; owned task construction for OSWorkerBench , 2026

   Abs HTML Code

   UI-Mate is an open-weight foundation GUI agent that pairs an environment-grounded training stack with in-context demonstration learning. Released alongside it is OSWorkerBench, a benchmark of 100 long-horizon office tasks spanning 41 applications, evaluating both general computer-use ability and the ability to follow a procedure supplied by a multimodal demonstration.

2. ICLR

   

   RobotArena ∞: Scalable Robot Benchmarking via Real-to-Sim Translation

   Yash Jangir, Yidi Zhang, Kashu Yamazaki, Chenyu Zhang, Kuan-Hsun Tu, and 4 more authors

   In International Conference on Learning Representations (ICLR), 2026

   Abs HTML

   RobotArena infinity shifts vision-language-action policy evaluation into large-scale simulated environments built automatically from real robot video. The pipeline converts demonstrations from widely used robot datasets into physics-consistent simulated counterparts using vision-language models, 2D-to-3D generative modeling and differentiable rendering, requiring only a single RGB image from a static-camera demonstration. Policies are scored by both VLM-guided automatic evaluation and scalable human preference judgments.

2025

1. CVPR

   

   SCAP: Transductive Test-Time Adaptation via Supportive Clique-Based Attribute Prompting

   Chenyu Zhang, Kunlun Xu, Zichen Liu, Yuxin Peng, and Jiahuan Zhou

   In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

   Abs

   SCAP preserves vision-language model accuracy under deployment distribution shift with no labels and no retraining, adapting the model dynamically based on cliques of visually similar test samples. Accuracy improves further as the test batch grows.