PeopleofSuperintelligence

Researcher

Hengzhi Pei

Hengzhi Pei is an applied scientist at Amazon who works on the trustworthiness and reliability of large language models. He coauthored DecodingTrust, a benchmark that tests GPT-4 and GPT-3.5 for toxicity, stereotype bias, privacy, machine ethics and fairness, which won an Outstanding Datasets and Benchmarks Paper award at NeurIPS 2023. Earlier papers covered multi-criteria Chinese word segmentation and the Secret Revealer model-inversion attack in 2019, and the RTSGAN time-series generator in 2021. At the University of Illinois at Urbana-Champaign, he completed a master's degree in computer science in 2023 with Bo Li as advisor, writing a thesis on TextGuard, a provable defense against backdoor attacks on text classification. His Amazon papers study silent data corruption in LLM training, knowledge distillation and LLM inference simulation.

Now at
Amazon
Badges

Timeline

  1. 2019

    Multi-criteria Chinese word segmentation

    In 2019, he coauthors with Xipeng Qiu, Hang Yan and Xuanjing Huang a unified Transformer-encoder model for multi-criteria Chinese word segmentation. The model segments simplified and traditional Chinese text according to a criterion token.

  2. 2019

    The Secret Revealer attack

    With Yuheng Zhang, Ruoxi Jia, Wenxiao Wang, Bo Li and Dawn Song, he coauthors "The Secret Revealer" in 2019. The paper presents a generative model-inversion attack that uses generative adversarial networks to invert deep neural networks and reconstruct face images from a face recognition model.

  3. 2021

    RTSGAN for real-world time series

    In 2021, he is first author of "Towards Generating Real-World Time Series Data" with Kan Ren, Yuqing Yang, Chang Liu, Tao Qin and Dongsheng Li. The paper proposes RTSGAN, a generative framework for long, variable-length time series with missing values.

  4. 2023

    Context for code language models

    At AAAI 2023, he is first author of "Better context makes better code language models," a case study on function call argument completion written with Jinman Zhao, Leonard Lausen, Sheng Zha and George Karypis. The work curates a new dataset of permissively licensed Python code.

  5. 2023

    A master's degree at Illinois

    He earns an M.S. in computer science at the University of Illinois at Urbana-Champaign in 2023, advised by Bo Li. His thesis presents TextGuard, a provable defense against backdoor attacks on text classification that splits training data into disjoint subsets and ensembles the resulting classifiers.

  6. 2023

    A NeurIPS Outstanding Paper Award

    DecodingTrust, which he coauthors with Boxin Wang, Weixin Chen, Bo Li and others, wins an Outstanding Datasets and Benchmarks Paper award at NeurIPS 2023 in the benchmark category. The benchmark evaluates GPT-4 and GPT-3.5 for toxicity, stereotype bias, adversarial robustness, privacy, machine ethics and fairness.

  7. 2025

    Silent data corruption and Amazon Nova

    At ACL 2025, he coauthors "Understanding silent data corruption in LLM training" with Jeffrey Ma, Leonard Lausen and George Karypis, comparing training on healthy production nodes with unhealthy nodes that show silent data corruption. In 2025, he is also among the authors of the technical report and model card for Amazon Nova, a family of foundation models.

  8. As of October 2026

    Applied scientist at Amazon

    He is an applied scientist at Amazon. His 2026 papers include an IEEE Micro study of silent data corruption across training characteristics, an ICLR 2026 paper on knowledge distillation for large language models through residual learning, and KernelSight-LM, a kernel-level LLM inference simulator.

Career

  1. NowAmazonApplied Scientist

Education

  1. 2023University of Illinois Urbana-ChampaignM.S.

From the biography.

  1. 16% of AI career moves in the 2020s went from a university to a companyIncludes Hengzhi Pei. We estimate that 16% of AI career moves in the 2020s went from a university to a company; 29% of moves starting at a university went to a company, and 76% ended the university role.
  2. Among employers with 30+ full-time cases, OpenAI leads primary technical role H-1B pay at $310,000, and Amazon, Google and Microsoft file 57% of all casesIncludes Amazon. Across 440,669 certified US H-1B filings from 44 employers (FY2020–FY2026), OpenAI leads primary technical role base pay at $310,000 and xAI leads overall base pay at $290,000.
  3. 20% of people at companies who shape AI also hold a university role, led by Google and Stanford UniversityIncludes Amazon. We estimate that 539 of the 6,766 people with a known current organization (8%) hold both a university role and a company role, including 20% of people at companies and 13% of people at universities.
  4. Google alumni have founded or led 133 other organizations, more than any other AI labIncludes Amazon. We estimate that alumni of 17 major AI labs went on to found or lead 350 other organizations, including 226 they founded, and we trace 472 career moves directly between 23 labs since 2010.

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 Hengzhi Pei yet. Send us a free-licensed photo.

  1. A Concise Model for Multi-Criteria Chinese Word Segmentation with Transformer Encoder arxiv.org. Checked 21 October 2026.
  2. The Secret Revealer: Generative Model-Inversion Attacks Against Deep Neural Networks arxiv.org. Checked 21 October 2026.
  3. Towards Generating Real-World Time Series Data arxiv.org. Checked 21 October 2026.
  4. Hengzhi Pei - Amazon Science amazon.science. Checked 21 October 2026.
  5. TextGuard: Provable defense against backdoor attacks on text classification | IDEALS ideals.illinois.edu. Checked 21 October 2026.
  6. Announcing the NeurIPS 2023 Paper Awards – NeurIPS Blog blog.neurips.cc. Checked 21 October 2026.
  7. NeurIPS 2023 Awards neurips.cc. Checked 21 October 2026.
  8. DecodingTrust: A Comprehensive Assessment of Trustworthiness in GPT Models arxiv.org. Checked 21 October 2026.
  9. The Amazon Nova Family of Models: Technical Report and Model Card arxiv.org. Checked 21 October 2026.
  10. KernelSight-LM: A Kernel-Level LLM Inference Simulator arxiv.org. Checked 21 October 2026.