Bo Zhang

B.Eng. in Automation, expected 2027

Beijing, China

bozhang23@mails.tsinghua.edu.cn

I am Bo Zhang (张博), an undergraduate student in the Department of Automation at Tsinghua University. I am interested in the interface between foundation models and real environments, with a focus on observation representations, action-space design, temporal context, and the translation of high-level intentions into reliable low-level actions.

Two recent papers approach this question from very different settings. In BAMI (CVPR 2026), I worked on training-free bias mitigation for GUI grounding, where an agent must connect visual perception to precise screen actions. EmbodiedAct (ICML 2026) studies LLM agents that drive scientific software like MATLAB — executing simulations, observing outcomes, and adjusting their plans mid-run. The settings are different, but the underlying question is how to make model behavior reliable once reasoning has to pass through an observation-action loop.

I complement paper-driven work with systems that expose real-world constraints. AICosmos explores multi-agent infrastructure for learning and teaching, where course context and feedback must be coordinated over time. Syll treats daily computing as an agent environment, with persistent context and user-facing actions. AEI4U brings course-specific AI assistance into analog electronics learning and won the Grand Prize at Tsinghua’s 43rd Challenge Cup. These projects keep my research tied to real users, real context, and real failure modes, exposing details that benchmarks often hide.

I am currently open to research collaborations on agent reliability, AI for science, and educational AI.

Featured projects

May 04, 2026	EmbodiedAct accepted to ICML 2026 — Grounding LLMs in Scientific Discovery via Embodied Actions. With Jinfeng Zhou, Yuxuan Chen, Jianing Yin, Minlie Huang, and Hongning Wang. code
Apr 18, 2026	Syll v0.2.0 released — recorded-workflow studio, layered memory workspace, and the desktop ghost are all live. project · code
Feb 28, 2026	BAMI accepted to CVPR 2026 — Training-Free Bias Mitigation in GUI Grounding. With Borui Zhang, Bo Wang, Wenzhao Zheng, and collaborators at Tsinghua and Lenovo Research. code
Dec 15, 2025	AICosmos — Tsinghua’s educational multi-agent platform — went public at aicosmos.ai.
Nov 15, 2025	AEI4U won the Grand Prize (特等奖) at Tsinghua’s 43rd Challenge Cup — AI Teaching-Agent Design Track, for an analog-electronics teaching assistant on Dify-THU.

selected publications

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}, }
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}, }

Featured projects

AICosmos

Syll

AEI4U

news

selected publications