Hi! I’m Chunxiao Li, a master’s student at the Beijing Normal University, majoring in Statistics, under the supervision of Prof. Chuanlong Xie. I am currently a research intern at Qiyuan National Lab, advised by Dr. Yao Zhu.
My research interests broadly include AI Safety, vision-language models (VLMs).
I am currently seeking PhD opportunities and aim to further explore AI safety, explainability, and the development of secure AI communities.
🔥 News
- 2025.03: 🎉 Our paper “Towards Annotation-Free Evaluation: KPA-Score for Human Keypoint Detection” was submitted to ICCV 2025! (Co-First author)
- 2025.03: 🎉 Our paper “Bridging the Gap Between Ideal and Real-world Evaluation: Benchmarking AI-Generated Image Detection in Challenging Scenarios” was submitted to ICCV 2025! (First author)
- 2025.03: 🎉 Our paper “Simplifying Debiasing: Random Feature Regularization as a Key Method” was submitted to ICCV 2025! (First author)
- 2025.3: 🎉 One paper on Noise Diffusion for Enhancing Semantic Faithfulness in Text-to-Image Synthesis accepted to CVPR 2025! (Second author)
- 2024.11: 🎉 Our work on An Efficient Framework for Enhancing Discriminative Models via Diffusion Techniques has been accepted by AAAI 2025! (Co-first author)
- 2024.11: 🎉 Our work on Prediction of iliac limb occlusion after endovascular aneurysm repair for abdominal aortic aneurysm by anatomical and near-wall hemodynamic characteristics combining numerical simulation and in vitro experiment has been accepted by Computer Methods and Programs in Biomedicine, IF=4.9. JCRQ1! (Co-first author)
📝 Publications
Towards Annotation-Free Evaluation: KPAScore for Human Keypoint Detection
Xiaoxiao Wang*, Chunxiao Li*, Peng Sun, Boming Miao, Yunjian Zhang, Yao Zhu
- We propose KPA-Score, a new annotation-free evaluation metric for keypoint detection based on vision-language models (VLMs). Our method simulates human judgment through binary response probability and correlates strongly with mAP, achieving 0.71 correlation without using ground-truth annotations.
Chunxiao Li*, Xiaoxiao Wang*, Meiling Li, Boming Miao, Peng Sun, Yunjian Zhang, Xiangyang Ji, Yao Zhu
- We introduce a new benchmark for AI-generated image detection that incorporates real-world distortions such as social media compression and re-digitization. Extensive experiments reveal performance gaps in existing detectors and VLMs. We also propose a robustness-aware in-context few-shot prompting method for improved detection accuracy.
Noise Diffusion for Enhancing Semantic Faithfulness in Text-to-Image Synthesis
Boming Miao, Chunxiao Li, Xiaoxiao Wang, Andi Zhang, Rui Sun, Zizhe Wang, Yao Zhu
- We propose a simple yet effective method to optimize the latent noise of diffusion models using VLM-guided question-answering, aiming to improve semantic faithfulness of generated images. Our method is model-agnostic, training-free, and improves both VQA and CLIP-based alignment scores.
An Efficient Framework for Enhancing Discriminative Models via Diffusion Techniques
Chunxiao Li*, Xiaoxiao Wang*, Boming Miao, Chuanlong Xie, Zizhe Wang, Yao Zhu
- Inspired by brain-like reasoning, we design a plug-and-play diffusion-enhanced framework for classification tasks. Low-confidence samples are reprocessed via conditional denoising paths, leading to consistent improvements across ImageNet-1K, CIFAR-10/100, and multiple robustness benchmarks.
🎖 Honors and Awards
-
2021.09 🏆 Second Prize (National), China Undergraduate Mathematical Contest in Modeling (CUMCM)
-
2022.12 🥉 Third Prize (National), China Undergraduate Statistical Modeling Competition
-
2021.12 🥉 Third Prize (National), China Undergraduate Statistical Modeling Competition
📖 Educations
- 2023.09 – 2025.06 (expected), 🎓 M.Sc. in Statistics, Beijing Normal University.
- 2019.09 – 2023.06, 🎓 B.Sc. in Information and Computing Science, University of Science and Technology Beijing.
🌱 Interests
I enjoy sports like Muay Thai, strength training, Brazilian Jiu-Jitsu, and more. I pursue a robust physique and a free spirit.