PeopleofSuperintelligence

Researcher

Jaeyeon Kim

Jaeyeon Kim is a computer science Ph.D. student at Harvard who studies generative models, including masked diffusion and one-step generation. While studying mathematics at Seoul National University from 2022 to 2024, he worked on optimization with Ernest K. Ryu and Asuman Ozdaglar before beginning his doctorate at Harvard in 2024 with Sitan Chen and Sham Kakade. In 2025, he was co-first author with Kulin Shah of "Train for the Worst, Plan for the Best: Understanding Token Ordering in Masked Diffusions," which won an Outstanding Paper Award at ICML 2025, and co-led the October 2025 PRISM paper on inference-time self-correction for masked diffusion models. He received a Kempner Institute graduate fellowship in 2026 and a 2026–2027 Jump Trading Fellowship, coauthored the Priming Models paper on one-step generation in October 2026, and is a member of the Sitan Chen and Kakade labs.

Timeline

  1. 2022–2024

    Mathematics and optimization at SNU

    From 2022 to 2024, he studies mathematics at Seoul National University, works as an undergraduate intern in Ernest K. Ryu's group, and works with Asuman Ozdaglar in 2023. In May 2023, he coauthors "Time-Reversed Dissipation Induces Duality Between Minimizing Gradient Norm and Function Value," which introduces H-duality between algorithms that minimize function values and those that minimize gradient magnitude. In April 2024, he coauthors "Optimal Acceleration for Minimax and Fixed-Point Problems is Not Unique," giving alternative acceleration mechanisms that achieve the same worst-case convergence rates as anchor-based methods.

  2. February 2025

    Token ordering in masked diffusions

    In February 2025, he and four coauthors post "Train for the Worst, Plan for the Best: Understanding Token Ordering in Masked Diffusions" on arXiv. The paper shows that choosing the token decoding order adaptively lets masked diffusion models sidestep hard subproblems, lifting Sudoku solving accuracy from under 7% to about 90%.

  3. 2025

    An ICML Outstanding Paper Award

    In July 2025, "Train for the Worst, Plan for the Best: Understanding Token Ordering in Masked Diffusions" wins an Outstanding Paper Award at ICML 2025, one of six papers honored out of roughly 12,000 submissions. Kim is its co-first author with Kulin Shah alongside Vasilis Kontonis, Sham Kakade and Sitan Chen, as a computer science Ph.D. student at Harvard SEAS.

  4. October 2025

    PRISM for masked diffusion

    In October 2025, he is a co-first author of "Fine-Tuning Masked Diffusion for Provable Self-Correction" with Seunggeun Kim and Taekyun Lee, alongside David Z. Pan, Hyeji Kim, Sham Kakade and Sitan Chen. The paper introduces PRISM, a lightweight fine-tuning approach that applies to any pretrained masked diffusion model and lets it detect and revise low-quality tokens at inference.

  5. 2026

    Kempner graduate fellowship

    In 2026, he is named among the recipients of the Kempner Institute's graduate fellowships at Harvard. He describes his research as advancing understanding of modern generative models and developing new generative modeling approaches for scientific and real-world applications.

  6. 2026–2027

    A Jump Trading Fellowship

    He is named a 2026–2027 Jump Trading Fellow from Harvard, listed under diffusion models and generative models.

  7. October 2026

    Priming Models for one-step generation

    In October 2026, he coauthors "Are We Overthinking One-Step Generation?" with Woobin Park, Yilun Du, Sitan Chen and Bohyung Han. The paper introduces Priming Models, a two-stage framework that trains a one-step generator on a reconstruction objective before feature-space distributional post-training.

  8. As of October 2026

    Ph.D. student at Harvard

    He begins his doctorate at Harvard in 2024 on an Ilju Scholarship, co-advised by Sitan Chen and Sham Kakade, and is a computer science Ph.D. student and a member of the Sitan Chen and Kakade labs. In September 2026, he presents "Attractor Models: One-Step Generative Modeling beyond Diffusion" at a Kempner Institute all-hands meeting. In October 2026, he serves as a junior organizer of the COLM 2026 workshop "Non-Autoregressive Language Models for Fast and Flexible Text Generation."

  1. More people shaping AI studied at Massachusetts Institute of Technology than at any other institutionIncludes Harvard University. We estimate where the people shaping AI studied and where their careers led. Massachusetts Institute of Technology leads with 467 former students, including 291 recorded doctorates.
  2. The median h-index is 48 at universities and 14 at frontier AI labsIncludes Harvard University. The median h-index is 48 at universities and 14 at frontier AI labs. We compare publishing careers, recent citations, dual appointments and job moves across 2,420 university researchers and 308 frontier-lab researchers.
  3. Companies supply the most prior-work hires at 7 of the 9 frontier AI labsIncludes Harvard University. Our estimates of the organizations where 550 people worked directly before joining 9 frontier AI labs across 620 career transitions, comparing the main feeder organizations per lab.
  4. 44% of AI career moves go from one university to anotherIncludes Harvard University. 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.

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 Jaeyeon Kim yet. Send us a free-licensed photo.

  1. Ernest K. Ryu ernestryu.com. Checked 22 October 2026.
  2. Jaeyeon Kim sites.google.com. Checked 22 October 2026.
  3. Time-Reversed Dissipation Induces Duality Between Minimizing Gradient Norm and Function Value arxiv.org. Checked 22 October 2026.
  4. Optimal Acceleration for Minimax and Fixed-Point Problems is Not Unique arxiv.org. Checked 22 October 2026.
  5. Train for the Worst, Plan for the Best: Understanding Token Ordering in Masked Diffusions arxiv.org. Checked 22 October 2026.
  6. Kempner Institute Collaborators Win Outstanding Paper Award at ICML 2025 kempnerinstitute.harvard.edu. Checked 22 October 2026.
  7. ICML 2025 Awards icml.cc. Checked 22 October 2026.
  8. Fine-Tuning Masked Diffusion for Provable Self-Correction arxiv.org. Checked 22 October 2026.
  9. Kempner Institute Announces Recipients of 2026 Graduate Fellowships kempnerinstitute.harvard.edu. Checked 22 October 2026.
  10. Fellowship Program jumptrading.com. Checked 22 October 2026.
  11. Are We Overthinking One-Step Generation? arxiv.org. Checked 22 October 2026.
  12. Sitan Chen sitanchen.com. Checked 22 October 2026.
  13. All Hands: Jaeyeon Kim, Lukas Fesser kempnerinstitute.harvard.edu. Checked 22 October 2026.
  14. Non-Autoregressive Language Models for Fast and Flexible Text Generation — COLM 2026 Workshop pengzhangzhi.github.io. Checked 22 October 2026.