01
Embodied Intelligence & Multimodal Agents
Perceive / reason / act
- Multimodal reasoning, GUI grounding, and embodied intelligent systems.
- Connecting model capabilities with real-world perception and interaction.
Researcher / Builder
Borui Zhang张博睿
Understanding intelligence. Building it into the world.
Postdoctoral Researcher/Tsinghua University
Co-founder & CEO/LoopEva
Leading research in the Embodied Intelligence Systems Group
About
I am a postdoctoral researcher at Tsinghua University, working with Prof. Jiwen Lu in the i-VisionGroup. I lead research in the Embodied Intelligence Systems Group, connecting advances in multimodal models with intelligent systems that perceive, reason, and act.
I am also the co-founder and CEO of LoopEva. Across research and entrepreneurship, I am interested in making intelligence more understandable, efficient, and useful in the physical world.
I received my Ph.D. in Control Science and Engineering from Tsinghua University in June 2026, advised by Prof. Jiwen Lu, and began my postdoctoral research in September 2026 as a Shuimu Scholar. Previously, I received a B.E. in Automation and a second B.A. in Economics from Tsinghua University in 2021.
Research
From the representations a model learns to the decisions a system makes, my research connects understanding, efficiency, and action.
02
Efficient Models & Visual Representation
Represent / train / deploy
- Visual tokenization, vector quantization, and generative model pretraining.
- Efficient training and deployment of multimodal foundation models.
03
Explainable Learning & Model Optimization
Interpret / diagnose / improve
- Mechanistic and axiomatic explanations of neural model behavior.
- Attribution, interpretable optimization, and reliable model decisions.
News
- Started postdoctoral research at Tsinghua University as a Shuimu Scholar, leading research in the Embodied Intelligence Systems Group.
- Received my Ph.D. from Tsinghua University.
- Our work on GUI grounding, BAMI, was accepted to CVPR 2026.
- Our work on explainable neural networks was accepted to TPAMI.
- Our work on fast Shapley value estimation was accepted to CJE.
Earlier updates
- 1 paper on autonomous driving accepted to ECCV 2024.
- 1 paper on salient object detection accepted to TGRS 2024.
- 2 papers on autonomous driving accepted to CVPR 2024.
- 1 paper on explainable deep networks accepted to ICLR 2024.
- 1 paper on explainable deep networks accepted to ICLR 2023.
- 1 paper on dynamic metric learning accepted to ECCV 2022.
- 1 paper on explainable metric learning accepted to CVPR 2022.
- 1 paper on deep metric learning accepted to ICCV 2021.
Selected Publications
* indicates equal contribution
Exploring Unified Perspective For Fast Shapley Value Estimation
This paper analyzes existing Shapley value estimators and proposes SimSHAP. Experiments validate that SimSHAP significantly accelerates the computation of accurate Shapley values.
BAMI: Training-Free Bias Mitigation in GUI Grounding
This paper analyzes grounding biases in MLLMs and proposes BAMI, a training-free framework. Experiments validate its state-of-the-art performance on ScreenSpot-Pro.
Alignment-Invertibility Regularization for Explainable Neural Networks
This journal extension of Bort introduces a plug-and-play optimizer enforcing boundedness and orthogonality constraints derived from alignment and invertibility to enhance explainability, with a data-aware extension DBort.
Path Choice Matters for Clear Attribution in Path Methods
We introduced the Concentration Principle and developed SAMP, an efficient model-agnostic interpreter incorporating infinitesimal constraint (IC) and momentum strategy (MS).
Bort: Towards Explainable Neural Networks with Bounded Orthogonal Constraint
This paper proposes Bort, an optimizer for improving model explainability with boundedness and orthogonality constraints, derived from model comprehensibility conditions.
Attributable Visual Similarity Learning
We propose AVSL, which employs a generalized similarity learning paradigm to represent the similarity between images with a graph for a more accurate and explainable measure.
Deep Relational Metric Learning
This paper proposes to adaptively learn an ensemble of features that characterizes an image from different aspects, employing a relational module to capture correlations among features.
Preprints
SFTok: Bridging the Performance Gap in Discrete Tokenizers
We investigate iterative approaches for constructing discrete tokenizers and propose SFTok. Analogous to the discrete diffusion paradigm, SFTok is well-suited for integration into Multimodal Large Models (MLLMs), facilitating the realization of a unified discrete diffusion framework.
Quantize-then-Rectify: Efficient VQ-VAE Training
To investigate the relationship between continuous and discrete tokenizers, we propose ReVQ. This method yields a high-performance VQ-VAE requiring only 40 GPU hours of training on a single RTX 4090.
Preventing Local Pitfalls in Vector Quantization via Optimal Transport
This study addresses the training instability of Vector-Quantized Networks (VQNs) by introducing OptVQ, a new method using the Sinkhorn algorithm for optimal transport. It achieves full codebook utilization (100%) and outperforms current VQNs.
Education & Appointments
Appointments
Sep. 2026 - Present
Postdoctoral Researcher · Shuimu Scholar
Prof. Jiwen Lu's laboratory. Leading research in the Embodied Intelligence Systems Group.
Current
Education
Sep. 2021 - Jun. 2026
Ph.D. in Control Science and Engineering
Tsinghua University · Department of Automation
Advised by Prof. Jiwen Lu. Visual model interpretability, multimodal model optimization, and visual representation learning.
Sep. 2017 - Jun. 2021
B.E. in Automation
Tsinghua University · Department of Automation
GPA: 3.81/4.0. Ranked 1st out of 181 students.
Sep. 2018 - Jun. 2021
B.A. in Economics · Second Degree
Tsinghua University · School of Economics and Management
A complementary foundation in economics and management.
Building at LoopEva
Co-founder & CEO / LoopEva
Robotics in a loop.
LoopEva connects robot capability, real-world deployment, and continuous improvement. As co-founder and CEO, I work at the intersection of research, technical direction, and building intelligent systems for the physical world.
Explore LoopEvaHonors & Awards
- 2026
- Shuimu Scholar, Tsinghua University
- 2026
- Outstanding Graduate, Beijing
- 2026
- Excellent Graduate, Tsinghua University
- 2025
- Outstanding Graduate Student CPC Member, Tsinghua University
- 2024
- 12.9 Student Counselor Award, Tsinghua University
- 2023
- National Scholarship
- 2022
- Outstanding Student Cadre Award, Tsinghua University
- 2022
- Excellent Teaching Assistant Award, Tsinghua University
- 2021
- Future Scholars Scholarship, Tsinghua University (rank: 1/181)
- 2020
- Changtong Scholarship
- 2019
- Jiang Nanxiang Scholarship
- 2018
- National Scholarship
Academic Services
- Conference reviewer
- CVPR (2022, 2024), ICCV (2023), ECCV, NeurIPS (2024), ICME (2022-2024), ACCV (2024), VCIP (2022)
- Journal reviewer
- IEEE Transactions on Image Processing
Teaching
Teaching Assistant · Numerical Analysis · Tsinghua University · 2021-2022