publications

2026

1. CVPR’26

   

   BAMI: Training-Free Bias Mitigation in GUI Grounding

   Borui Zhang^*, Bo Zhang^*, Bo Wang^*, Wenzhao Zheng, Yuhao Cheng, and 4 more authors

   In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026

   Abs Bib Code

   GUI grounding is a critical capability for enabling GUI agents to execute tasks such as clicking and dragging. Using our Masked Prediction Distribution (MPD) attribution method, we identify two primary error sources: precision bias from high image resolution and ambiguity bias from intricate interface elements. We introduce Bias-Aware Manipulation Inference (BAMI), which combines coarse-to-fine focus and candidate selection to mitigate these biases in a training-free setting. BAMI lifts TianXi-Action-7B accuracy on ScreenSpot-Pro from 51.9% to 57.8% and consistently improves multiple grounding backbones across diverse benchmarks.

   @inproceedings{zhang2026bami,
     title = {{BAMI}: Training-Free Bias Mitigation in {GUI} Grounding},
     author = {Zhang, Borui and Zhang, Bo and Wang, Bo and Zheng, Wenzhao and Cheng, Yuhao and Tang, Liang and Yan, Yiqiang and Zhou, Jie and Lu, Jiwen},
     booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
     year = {2026},
   }

2. ICML’26

   

   Grounding LLMs in Scientific Discovery via Embodied Actions

   Bo Zhang^*, Jinfeng Zhou^*, Yuxuan Chen, Jianing Yin, Minlie Huang, and 1 more author

   In Proceedings of the International Conference on Machine Learning (ICML), 2026

   Abs Bib Code

   Large Language Models have shown significant potential in scientific discovery but struggle to bridge theoretical reasoning and verifiable physical simulation. Existing solutions operate in a passive "execute-then-response" loop and lack runtime perception, blinding agents to transient anomalies such as numerical instability. We propose EmbodiedAct, a framework that transforms established scientific software into active embodied agents by grounding LLMs in embodied actions with a tight perception-execution loop. Instantiated within MATLAB, EmbodiedAct achieves SOTA performance on complex engineering design and scientific modeling, with strong reliability and stability over long-horizon simulations.

   @inproceedings{zhang2026embodiedact,
     title = {Grounding {LLMs} in Scientific Discovery via Embodied Actions},
     author = {Zhang, Bo and Zhou, Jinfeng and Chen, Yuxuan and Yin, Jianing and Huang, Minlie and Wang, Hongning},
     booktitle = {Proceedings of the International Conference on Machine Learning (ICML)},
     year = {2026},
   }