PeopleofSuperintelligence

Researcher

Nenghai Yu

Also known as 俞能海, Yu Nenghai

Nenghai Yu is a professor at the University of Science and Technology of China (USTC) who works on multimedia security, information hiding and the safety of AI systems. He earned a Tsinghua master's degree in 1992 and a USTC doctorate in 2004, and by 2023 he was Executive Dean of USTC's School of Cyber Security and Director of its Information Processing Center. His coauthored work includes Dual Learning for Machine Translation in 2016, Multi-attentional Deepfake Detection and the CSWin Transformer in 2021, and the OPERA method against hallucination in multimodal language models at CVPR 2024. In 2026 he coauthored a paper on the limits of LLM safeguards and PURA, a watermarking method for AI-generated text. He is a professor in USTC's School of Cyber Security.

Timeline

  1. 1987–2004

    Tsinghua master's and a USTC doctorate

    He receives a B.S. from Nanjing University of Posts and Telecommunications in 1987 and an M.E. from Tsinghua University in 1992. In 2004, he earns a Ph.D. from the Department of Electronic Engineering and Information Science at USTC in Hefei.

  2. 2016

    Dual learning for translation

    In 2016, he coauthors Dual Learning for Machine Translation, a NIPS 2016 poster with Di He, Yingce Xia, Tao Qin, Liwei Wang, Tie-Yan Liu and Wei-Ying Ma. One model translates in each direction, and the two models teach each other through reinforcement learning, so a system can learn from unlabeled data.

  3. 2021

    Multi-attentional deepfake detection

    In 2021, he coauthors Multi-attentional Deepfake Detection with Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei and Weiming Zhang, published at CVPR 2021. The network treats deepfake detection as fine-grained classification, with multiple spatial attention heads, a textural feature enhancement block and attention-guided feature aggregation.

  4. 2021

    The CSWin Transformer backbone

    In July 2021, he coauthors CSWin Transformer, a vision backbone built on cross-shaped window self-attention, with Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen and Baining Guo. It reaches 85.4% Top-1 accuracy on ImageNet-1K without extra training data, 1.2 points above the Swin Transformer backbone.

  5. By 2023

    Executive dean at USTC

    By 2023, he is a professor in USTC's School of Cyber Science and Technology and School of Information Science and Technology, Executive Dean of the School of Cyber Security, and Director of the Information Processing Center. He is the author or coauthor of more than 130 papers, with research interests in multimedia security, information hiding and quantum networking.

  6. 2023–2024

    OPERA against MLLM hallucination

    In November 2023, he coauthors OPERA with Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin and Weiming Zhang, and it appears at CVPR 2024. OPERA is a decoding method for multimodal large language models that penalizes over-trust in summary tokens during beam search and rolls back to reallocate token choices, which reduces hallucination without extra data or training.

  7. July–August 2026

    Watermarking and LLM safeguards

    In July 2026, he coauthors a paper with Pingyu Wu, Lingyao Zhu and Weiming Zhang showing that Useful Capability, Reliable Safety and Open Access cannot coexist in LLM safeguards built on copyable context. In August 2026, he coauthors PURA, a multi-bit watermarking method for attributing AI-generated text that preserves the base generation distribution exactly and keeps recovery stable under post-editing.

  8. As of October 2026

    Professor at USTC

    He is a professor at USTC in the School of Cyber Security, the affiliation he gives for PseudoFix, a 2025 ASE paper on refactoring distorted structures in decompiled C pseudocode.

Career

  1. NowUniversity of Science and Technology of ChinaProfessor

Education

  1. 2004University of Science and Technology of ChinaPh.D.
  2. 1992Tsinghua UniversityM.E.
  3. 1987Nanjing University of Posts and TelecommunicationsB.S.

From the biography.

  1. 84% of people shaping AI with recorded undergraduate and doctoral schools changed institutions for their doctorateIncludes University of Science and Technology of China. 84% of people shaping AI with recorded undergraduate and doctoral schools changed institutions for their doctorate. Across 2,116 people with both an undergraduate degree and a doctorate recorded, we trace which undergraduate schools feed which doctoral programs, how often students stay at the same school, and how paths differ by birth-country region.
  2. 44% of AI career moves go from one university to anotherIncludes University of Science and Technology of China. We estimate how 26,439 AI career moves link universities, companies, labs, nonprofits and government research centers, with decade counts, net flows and the top paths.
  3. 11% of AI doctorate holders from 2010 to 2016 founded or led a company or lab by 10 years after graduationIncludes University of Science and Technology of China. We estimate how often AI doctorate holders founded or led a company or lab by 10 years after graduation, how waiting times compare between founders and top jobs, and which universities trained the most.
  4. The Massachusetts Institute of Technology trained or employed 44 of the 187 AI major-prize winners, more than any other universityIncludes University of Science and Technology of China. The Massachusetts Institute of Technology trained or employed 44 of the 187 AI major-prize winners, more than any other university. We compare doctorate and 5+ year employment links, career timing and current workplaces for 187 major-prize and 1,111 best-paper winners.

Colleagues

People named in this biography, and people with dated positions at the same organization in the same years.

Sources and credits

Photo: We have no published photo of Nenghai Yu yet. Send us a free-licensed photo.

  1. Entanglement-Assisted Quantum Networks: Mechanics, Enabling Technologies, Challenges, and Research Directions arxiv.org. Checked 22 October 2026.
  2. NIPS Poster Dual Learning for Machine Translation neurips.cc. Checked 22 October 2026.
  3. Multi-Attentional Deepfake Detection (CVPR 2021 Open Access Repository) openaccess.thecvf.com. Checked 22 October 2026.
  4. Multi-attentional Deepfake Detection - Microsoft Research microsoft.com. Checked 22 October 2026.
  5. CSWin Transformer: A General Vision Transformer Backbone with Cross-Shaped Windows arxiv.org. Checked 22 October 2026.
  6. Nenghai Yu | DeepAI deepai.org. Checked 22 October 2026.
  7. OPERA: Alleviating Hallucination in Multi-Modal Large Language Models via Over-Trust Penalty and Retrospection-Allocation arxiv.org. Checked 22 October 2026.
  8. PURA: Provably Unbiased and Robust Multi-Bit Watermarking for AI-Generated Text Attribution arxiv.org. Checked 22 October 2026.
  9. Safeguards Based on Copyable Context Cannot Provide Reliable Safety for LLMs arxiv.org. Checked 22 October 2026.
  10. Nenghai Yu - ASE 2025 conf.researchr.org. Checked 22 October 2026.