Bingchen Zhao

I am a Ph.D student at the University of Edinburgh, supervised by Dr Oisin Mac Aodha.
I am interested in Open Endedness, Self-Improving/Supervised Agents, and Interpretable AI.
Please feel free to drop me an email if you are interested in what I do and looking for possible collaborations.

Contact: zhaobc.gm@gmail.com

News

06/2025	Semanticist accepted by ICCV 2025!
02/2024	One paper accepted by CVPR 2024, my first CVPR paper!
01/2024	One paper accepted by ICLR 2024 as Spotlight!, see you in Vienna!
03/2023	Our 2nd OOD-CV workshop is accepted at ICCV, stay tuned for more details, see you in Paris!
10/2022	Recognised as a Top Reviewer for NeurIPS 2022!.
07/2022	Two papers accepted by ECCV 2022 with one selected as Oral!.
04/2022	We are organizing a workshop at ECCV 2022, check it out here.
09/2021	One paper accepted into NeurIPS 2021!
07/2021	One paper accepted into ICCV 2021 as Oral!
09/2019 - 05/2022	I am working as a teaching assistant for Prof. Yin Wang's Deep Learning Course at Tongji University.

Publications

Preprints / Recent Papers

“Principal Components” Enable A New Language of Images
Xin Wen*, Bingchen Zhao*, Ismail Elezi, Jiankang Deng, Xiaojuan Qi
arXiv / Code
ICCV 2025
TL;DR: A tokenizer that decouples semantic and spectrum information with variable length tokens!

Conference Papers

	What If We Recaption Billions of Web Images with LLaMA-3? Xianhang Li, Haoqin Tu, Mude Hui, Zeyu Wang, Bingchen Zhao*, Junfei Xiao, Sucheng Ren, Jieru Mei, Qing Liu, Huangjie Zheng, Yuyin Zhou, Cihang Xie arXiv / Code ICML 2025 TL;DR: We generated the caption of for the DataComp-1B dataset and demonstrate an improved performance on various tasks.
	HQ-Edit: A High-Quality Dataset for Instruction-based Image Editing Mude Hui, Siwei Yang, Bingchen Zhao, Yichun Shi, Heng Wang, Peng Wang, Cihang Xie, Yuyin Zhou arXiv / Code ECCV 2024 TL;DR: A high quality dataset for image editing is proposed.
	Eagle and Finch: RWKV with Matrix-Valued States and Dynamic Recurrence RWKV Project arXiv / Code COLM 2024 TL;DR: An RNN-based LLM in the era of Transformers! And now being deployed to half a billion systems worldwide!
	What Makes CLIP More Robust to Long-Tailed Pre-Training Data? A Controlled Study for Transferable Insights Xin Wen, Bingchen Zhao, Yilun Chen, Jiangmiao Pang, Xiaojuan Qi arXiv / Code NeurIPS 2024 TL;DR: We find CLIP to be relatively robust to pre-training data imbalance, design and conduct controlled experiments to identify the underlying mechanisms and provide insights for recognition and SSL models..
	Sight Beyond Text: Multi-Modal Training Enhances LLMs in Truthfulness and Ethics Haoqin Tu, Bingchen Zhao, Chen Wei, Cihang Xie arXiv / Code TMLR TL;DR: We present a surprising finding that multi-modal tuning improves LLMs in truthfulness and ethical behaviors.
	Vision Learners Meet Web Image-Text Pairs Bingchen Zhao, Quan Cui, Hao Wu, Osamu Yoshie, Cheng Yang, Oisin Mac Aodha arXiv / Website TMLR 2024 TL;DR: We present a visual representation pre-training method for scalable web image-text data and it achieves state-of-the-art performance on various tasks with promising scaling behavior.
	Labeled Data Selection for Category Discovery Bingchen Zhao, Nico Lang, Serge Belongie, Oisin Mac Aodha arXiv ECCV 2024 TL;DR: What labeled data to use in category discovery matters, and we present two methods for selecting the right labeled data to use.
	PromptCCD: Learning Gaussian Mixture Prompt Pool for Continual Category Discovery Fernando Cendra, Bingchen Zhao, Kai Han arXiv / Code ECCV 2024 TL;DR: We present a framework for the Continual Category Discovery (CCD) task, and demonstrate state-of-the-art performance.
	How Many Unicorns Are in This Image? A Safety Evaluation Benchmark for Vision LLMs Haoqin Tu, Chenhang Cui, Zijun Wang*, Yiyang Zhou, Bingchen Zhao, Junlin Han, Wangchunshu Zhou, Huaxiu Yao, Cihang Xie arXiv / Code ECCV 2024 TL;DR: We provide the first comprehensive safety benchmark for VLLMs, including OOD scenarios and Redteaming attacks.
	Fool Your (Vision and) Language Model With Embarrassingly Simple Permutations Yongshuo Zong, Tingyang Yu, Bingchen Zhao, Ruchika Chavhan, Timothy Hospedales arXiv / Code ICML 2024 TL;DR: We observed that a surprisingly simple answer permutation degrades performance of LLMs.
	What If the TV Was Off? Examining Counterfactual Reasoning Abilities of Multi-modal Language Models Letian Zhang, Xiaotong Zhai, Zhongkai Zhao, Yongshuo Zong, Xin Wen, Bingchen Zhao arXiv / Code / Page CVPR 2024 TL;DR: Vision-Language Models cannot handle counterfactual questions very well.
	Tuning LayerNorm in Attention: Towards Efficient Multi-Modal LLM Finetuning Bingchen Zhao, Haoqin Tu, Chen Wei, Jieru Mei, Cihang Xie arXiv / Code ICLR 2024 Spotlight TL;DR: We explored a parameter efficient fine-tuning method that reduces the parameters by 10x than the LoRA fine-tuning while achieving similar performance.
	Incremental Generalized Category Discovery Bingchen Zhao, Oisin Mac Aodha arXiv Code / Page ICCV 2023 TL;DR: We propose a new setting named incremental generalized category discovery that requires the model to learn to discover novel categories at each new incremental stages, a simple baseline model is also proposed based on non-parametric classifiers.
	Learning Semi-supervised Gaussian Mixture Models for Generalized Category Discovery Bingchen Zhao, Xin Wen, Kai Han arXiv / Code ICCV 2023 TL;DR: We tackle GCD without knowing the class number, propose a semi-supervised variant of GMM with stochastic splitting and merging to dynamically determine prototypes, and leverage PCL for representation learning on partially labelled data.
	Parametric Classification for Generalized Category Discovery: A Baseline Study Xin Wen, Bingchen Zhao, Xiaojuan Qi arXiv / Code ICCV 2023 TL;DR: We revisit the reason that makes previous parametric classifiers fail to recognise new classes for GCD, identify the prediction biases between and within seen and novel classes as the key issue, and propose a simple yet strong framework that addresses these limitations and achieves state-of-the-art performance in this field.
	XCon: Learning with Experts for Fine-grained Category Discovery Yixin Fei, Zhongkai Zhao, Siwei Yang, Bingchen Zhao arXiv / Slides / Code BMVC 2022 Oral (34/770=4.4%) TL;DR: Learning to do category discovery within a fine-grained dataset is challenging, we present a method that learn to do that by partition the dataset into k sub-groups, and show improved performance on several fine-grained datasets.
	Self-Supervised Visual Representation Learning with Semantic Grouping Xin Wen, Bingchen Zhao, Anlin Zheng, Xiangyu Zhang, Xiaojuan Qi arXiv / Website / Code NeurIPS 2022 TL;DR: Our model can do scene decomposition and representation learning at the same time and shows strong generalization ability pretrained on scene-centric data.
	OOD-CV: A Benchmark for Robustness to Out-of-Distribution Shifts of Individual Nuisances in Natural Images Bingchen Zhao, Shaozuo Yu, Wufei Ma, Mingxin Yu, Shenxiao Mei, Angtian Wang, Ju He, Alan Yuille, Adam Kortylewski. arXiv / Website / Download / Slides / TPAMI Verison ECCV 2022 Oral (158/5803=2.7%) TPAMI 2024 TL;DR: We collected a dataset where we have the control over the individual OOD attribute in the test examples.
	Discriminability-Transferability Trade-Off: An Information-Theoretic Perspective Quan Cui, Bingchen Zhao, Zhao-Min Chen, Borui Zhao, Renjie Song, Jiajun Liang, Boyan Zhou, Osamu Yoshie. arXiv / Code / Slides ECCV 2022 TL;DR: We study the transferability and the discriminability of deep representations and found a trade-off between these two properties.
	Novel Visual Category Discovery with Dual Ranking Statistics and Mutual Knowledge Distillation Bingchen Zhao, Kai Han. arXiv / Code / Slides NeurIPS 2021 TL;DR: We extend novel category discovery to discover fine-grained classes by leverging information from image parts.
	Improving Contrastive Learning by Visualizing Feature Transformation Rui Zhu, Bingchen Zhao, Jingen Liu, Zhenglong Sun, Chang Wen Chen. arXiv / Code / Slides ICCV 2021 Oral (210/6236=3.4%) TL;DR: We explore the training dynamics of self-supervised contrastive learning, and proposed two simple method for improving the performance of the model.
	Temporal Context Aggregation for Video Retrieval with Contrastive Learning Jie Shao, Xin Wen, Bingchen Zhao, Xiangyang Xue. arXiv / Code / Slides WACV 2021 TL;DR: Video retrieval methods can be improved by modeling long-range temporal information with transformer and contrastive learning.

Workshop Papers

One Venue, Two Conferences: The Separation of Chinese and American Citation Networks
Bingchen Zhao*, Yuling Gu*, Jessica Zosa Forde, Naomi Saphra
arXiv
NeurIPS 2022 AI Cultures Workshop
TL;DR: At NeurIPS, American and Chinese institutions cite papers from each other's regions substantially less than they cite endogamously. We build a citation graph to quantify this divide, compare it to European connectivity, and discuss the causes and consequences of the separation.

Distilling Visual Priors from Self-Supervised Learning
Bingchen Zhao, Xin Wen
arXiv / Code / Slides
ECCV 2020 VIPriors Workshop
TL;DR: Learning a model self-supervisedly and then do self-distillation helps in the data-deficient domain.

Awards

2022	Top-Reviewer for NeurIPS 2022.
2020	First-place in the FGVC7 workshop iWildcam challenge track.
2020	Second-place in the ECCV 2020 VIPrior workshop image classification challenge track.
2020	Best Undergraduate Prize in the NeurIPS 2020 SpaceNet 7 challenge.
2016	Bronze medal in the Asia-Pacific Informatics Olympiad.
2015	First Prize in the National Olympiad in Informatics in Provinces.

Professional Services

I have been a reviewer for ICLR, NeurIPS, CVPR, ICCV, ECCV, WACV, FGVC, and SIGSPATIAL.