I received my Ph.D. degree from the Department of Automation, Tsinghua University in 2026, where I was advised by Prof. Jiwen Lu at the i-VisionGroup. I will join Prof. Jiwen Lu's team as a postdoctoral researcher, where I will coordinate two research groups focusing on on-device agents and embodied brain. I received my B.E. degree from the Department of Automation and a second B.A. degree from the School of Economics and Management at Tsinghua University in 2021. My research aims to build interpretable, efficient, and embodied large-scale intelligent systems.
My current research focuses on the principles, systems, and applications of large-scale intelligent models. In particular, I study how to make large models more interpretable, lightweight, and agentic, with an emphasis on connecting theoretical understanding with practical deployment in multimodal and embodied scenarios.
* indicates equal contribution
arXiv, 2025
Paper/ Code/ Project Page/ HF Demo/ 中文解读
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.
arXiv, 2025
Paper/ Code/ Project Page/ HF Demo/ 中文解读
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.
arXiv, 2024
Paper/ Code/ Project Page/ HF Demo/ 中文解读
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.
CJE, 2026
This paper analyzes existing Shapley value estimators and proposes SimSHAP. Experiments validate that SimSHAP significantly accelerates the computation of accurate Shapley values.
CVPR, 2026
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.
TPAMI, 2026
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.
ICLR, 2024
We introduced the Concentration Principle and developed SAMP, an efficient model-agnostic interpreter incorporating infinitesimal constraint (IC) and momentum strategy (MS).
ICLR, 2023
This paper proposes Bort, an optimizer for improving model explainability with boundedness and orthogonality constraints, derived from model comprehensibility conditions.
CVPR, 2022
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.
ICCV, 2021
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.
| 2025 | Tsinghua University Outstanding Graduate Student CPC Member |
|---|---|
| 2024 | 12·9 Student Counselor Award |
| 2023 | National Scholarship (Highest national honor for students) |
| 2022 | Tsinghua Outstanding Student Cadre Award |
| 2022 | Tsinghua Excellent Teaching Assistant Award |
| 2021 | Tsinghua Future Scholars Scholarship (Rank: 1/181) |
| 2020 | Changtong Scholarship (Highest dept. honor) |
| 2019 | Jiang Nanxiang Scholarship (Highest junior honor) |
| 2018 | National Scholarship |
| Conference Reviewer | CVPR (2022, 2024), ICCV (2023), ECCV, NeurIPS 2024, ICME (2022-2024), ACCV 2024, VCIP 2022 |
|---|---|
| Journal Reviewer | IEEE Transactions on Image Processing (T-IP) |