Quanquan Gu

wiki
, ,
1

Machine learning theorist and AI research leader who simultaneously advanced protein structure prediction at the frontier of AI for Science and co-led LLM pretraining and scaling at ByteDance Seed.
Born c. 1984, China
Nationality Chinese-American
Current Institution(s) University of California, Los Angeles — Department of Computer Science (Associate Professor with Tenure, on leave); ByteDance Research (Research Scientist, July 2023–June 2026, departed)
Research Areas Nonconvex Optimization, Deep Learning Theory, Reinforcement Learning, Large Language Models, AI for Science (Protein Design & Structure Prediction), Diffusion Models
Doctoral Advisor Jiawei Han
Doctoral Thesis Online and Active Learning of Big Networks: Theory and Algorithms (University of Illinois at Urbana-Champaign, 2014)
Website web.cs.ucla.edu/~qgu
X / Twitter @QuanquanGu
Google Scholar Quanquan Gu

Overview

Quanquan Gu is a Chinese-American machine learning researcher and tenured Associate Professor of Computer Science at UCLA, known for rigorous theoretical contributions to deep learning, nonconvex optimization, and reinforcement learning. From July 2023 to June 2026 he held a concurrent research scientist role at ByteDance Seed, where he led two distinct research fronts: first building one of the most competitive AI for protein science stacks in industry, then pivoting to co-found the LLM optimization and scaling team that underpinned the development of Seed 2.0. His parallel command of ML theory, biological AI, and large-scale pretraining—fields that rarely overlap in a single researcher—has drawn sustained attention in both academia and industry. He departed ByteDance in June 2026, signing off with the phrase “The best model is yet to come. Scaling continues,” without disclosing his next destination.

Early Life & Education

Gu completed his undergraduate studies at Tsinghua University, earning a Bachelor of Engineering in Automation (2003–2007), followed by a Master of Science in Control Science and Engineering from the same institution (2007–2010). He then moved to the United States to pursue a doctorate at the University of Illinois at Urbana-Champaign (UIUC) under the supervision of Jiawei Han, one of the most cited researchers in data mining and knowledge discovery. His dissertation, Online and Active Learning of Big Networks: Theory and Algorithms, addressed the challenge of efficiently learning from massive, rapidly evolving information networks, combining online learning algorithms with active learning strategies. He completed his PhD in 2014 and received both the UIUC Computer Science Department Fellowship (2010) and the IBM PhD Fellowship (2013–2014) during his doctoral studies. Following graduation, he spent a year as a Postdoctoral Research Associate in the Department of Operations Research and Financial Engineering at Princeton University (2014–2015).

Career

University of Virginia (2015–2018)

Gu joined UVa as a tenure-track Assistant Professor, initially in the Department of Systems and Information Engineering (2015–2017), with a concurrent appointment in the Department of Computer Science added from 2016 until his departure in 2018. During this period he established his research agenda in nonconvex optimization and the theoretical foundations of deep learning, publishing early work on stochastic variance reduction methods and statistical learning theory for overparameterized models. He received the NSF CAREER Award (2017) and the UVa SEAS Research Innovation Award (2017) while at UVa, and secured multiple NSF grants on machine learning with privacy and high-dimensional data.

University of California, Los Angeles, Department of Computer Science (2018–present)

Gu joined UCLA as a tenure-track Assistant Professor in July 2018, was promoted to Associate Professor with Tenure in July 2022, and leads the UCLA Artificial General Intelligence Lab. His research group has produced a substantial body of work across four interconnected areas.

In optimization and deep learning theory, the lab derived sharp generalization bounds for stochastic gradient descent in overparameterized networks, characterized benign overfitting in two-layer convolutional neural networks (NeurIPS 2022, Oral), and analyzed the implicit regularization of SGD across a series of papers. The MARS optimizer (ICML 2025) introduced a new variance-reduction framework for training large models, and Tensor Product Attention (NeurIPS 2025, Spotlight) proposed a new attention mechanism architecture.

In reinforcement learning, the group established near-minimax optimal sample complexity results for linear Markov decision processes and linear mixture MDPs (COLT 2021, ICML 2021, ICML 2023), contributing to a rigorous theory of provably efficient RL with function approximation.

In language model alignment and fine-tuning, Gu co-developed Self-Play Fine-Tuning (SPIN, ICML 2024), a method enabling a language model to improve iteratively by playing against its own previous-iteration outputs without requiring additional human annotations, and extended the framework to diffusion models for text-to-image generation (NeurIPS 2024). Self-Play Preference Optimization (SPPO, ICLR 2025) refined this into a principled alignment approach.

In AI for science, UCLA-based work on structure-based drug design (DecompDiff, ICLR 2024; DecompOpt, ICLR 2024) and protein conformation modeling (ICML 2024) directly prefigured and informed the industrial-scale biology research Gu would later lead at ByteDance.

During the COVID-19 pandemic (2020–2021), his team developed the UCLA SuEIR epidemiological model, which was used by the US Centers for Disease Control and Prevention for weekly forecasts of cumulative deaths and hospitalizations, with results published in PNAS (2022, 2023).

Gu also held two extended research residencies at the Simons Institute for the Theory of Computing: as a Research Fellow in the Foundations of Deep Learning Summer Program at Berkeley (May–August 2019), and as a Long-term Participant in the Theory of Reinforcement Learning Program (August–December 2020). He was additionally a Short-term Visitor at the Institute for Advanced Study in Princeton during the Special Year on Optimization, Statistics, and Theoretical Machine Learning (October–November 2019).

ByteDance Research (July 2023–June 2026)

Gu joined ByteDance Research in July 2023 to lead the company’s AI for Drug Discovery initiative, concurrently maintaining his UCLA faculty position on leave. Over roughly eighteen months his team produced three major outputs in computational biology.

SeedFold (arXiv: 2512.24354, December 2025) is a biomolecular structure prediction system that scales the AlphaFold 3 architecture along three axes—model width, architectural efficiency via linear triangular attention, and data volume through large-scale distillation to 26.5 million training samples. Benchmarked on FoldBench, SeedFold surpassed AlphaFold 3 on most protein-related tasks including monomer prediction, protein-protein docking, antibody-antigen complexes, and protein-ligand binding.

SeedProteo (arXiv: 2512.24192, February 2026) is a diffusion-based de novo all-atom protein binder design system that repurposes the SeedFold folding architecture as a generative framework, integrating self-conditioning features to guide the design process. In unconditional generation it handles sequences up to 1,000 residues, and in binder design benchmarks it surpassed AlphaProteo, RFDiffusion, Chai-2, BinderCraft, and BoltzGen. In vitro validation on PD-L1 and SC2RBD targets confirmed experimentally expressed and binding binders.

DPLM / DPLM-2 is a family of diffusion protein language models (ICML 2024; ICLR 2025; extended in ICML 2025) that unify protein representation learning, unconditional generation, and any-to-any conditional generation—including folding, inverse folding, and motif scaffolding—under a single discrete-diffusion pre-training objective. DPLM-2 extends the framework to jointly model sequence and 3D structure within a single pretrained backbone.

In early 2025, following DeepSeek’s high-profile releases that catalyzed intensified LLM competition in China, Gu pivoted to join ByteDance’s frontier model effort, founding the LLM optimization and scaling team within the Seed pretraining group. This team built a scalable pretraining stack from scratch and directly contributed to the development and release of Seed 2.0—ByteDance’s flagship frontier model—roughly four months before Gu’s departure. ByteDance’s AI assistant Doubao, which integrates Seed 2.0 capabilities, had reached approximately 336 million monthly active users by mid-2026.

Gu announced his departure from ByteDance Seed on June 3, 2026, thanking colleagues for “an incredibly rewarding journey” without disclosing his next move.

Key Contributions

  • SeedFold — The first industrially deployed biomolecular structure prediction system to surpass AlphaFold 3 across a broad range of FoldBench tasks, combining architecture scaling, linear attention, and large-scale data distillation to 26.5M samples.
  • SeedProteo — State-of-the-art de novo all-atom protein binder design model, outperforming all major open-source competitors in in-silico benchmarks with experimental in-vitro validation; publicly deployed at seedfold.io.
  • DPLM series — A family of diffusion protein language models (ICML 2024, ICLR 2025) unifying representation learning and conditional/unconditional generation; DPLM-2 jointly models sequence and 3D structure within a single pretrained backbone.
  • Seed 2.0 pretraining stack — Co-led the LLM optimization and scaling team that built ByteDance’s scalable frontier pretraining infrastructure, directly powering Seed 2.0 and subsequent models.
  • Self-Play Fine-Tuning (SPIN) — A language model self-improvement framework (ICML 2024) enabling iterative improvement without additional human annotation; extended to diffusion models (NeurIPS 2024) and preference optimization (SPPO, ICLR 2025).
  • Benign overfitting theory — A series of papers characterizing when overparameterized neural networks trained by SGD generalize despite interpolating training data (NeurIPS 2022 Oral; COLT 2021; JMLR 2024), providing rigorous grounding for a widespread empirical phenomenon in modern deep learning.
  • Provably efficient reinforcement learning — Established near-minimax optimal sample complexity bounds for linear MDPs and linear mixture MDPs (COLT 2021, ICML 2021, ICML 2023), placing RL algorithms on rigorous theoretical foundations.
  • UCLA SuEIR model — COVID-19 epidemiological model used by the CDC for national forecasts of deaths and hospitalizations (2020–2021), with results contributing to two PNAS publications (2022, 2023).
  • EurekaClaw — A local-first AI research agent (eurekaclaw.ai) that automates the pipeline from ideation through proof, experiment, and paper generation, embodying Gu’s broader vision of AI-augmented scientific discovery.

Awards & Recognition

  • WSDM Test of Time Paper Award (2024) — Recognizing long-term impact of early research on information networks and learning.
  • Alfred P. Sloan Research Fellowship (2022) — Awarded by the Alfred P. Sloan Foundation for outstanding early-career contributions to machine learning and AI.
  • JP Morgan Faculty Research Award (2022) — Corporate recognition of applied ML research impact.
  • AWS Machine Learning Research Award (2020) — Amazon Web Services recognition of machine learning research contributions.
  • IJCAI Early Career Talk (2020) — Selected to deliver an early-career invited talk at the International Joint Conference on Artificial Intelligence.
  • Simons Berkeley Research Fellowship (2019) — Fellowship at the Simons Institute for the Theory of Computing, Foundations of Deep Learning Summer Program.
  • Adobe Data Science Research Award (2018) and Salesforce Deep Learning Research Award (2018) — Dual industry recognition in the same year.
  • NSF CAREER Award (2017) — The US National Science Foundation’s principal early-career award, supporting his research program on nonconvex statistical optimization.
  • UVa SEAS Research Innovation Award (2017) — University of Virginia School of Engineering recognition of research innovation.
  • Yahoo! Academic Career Enhancement Award (2015) — Awarded at the outset of his faculty career.
  • IBM PhD Fellowship (2013–2014) — Competitive IBM fellowship supporting doctoral research.
  • Best Master Thesis Award, Tsinghua University (2010) — Departmental recognition of outstanding master’s thesis work.
  • UIUC Computer Science Department Fellowship (2010) — Fellowship supporting entry into the doctoral program at UIUC.

Key Relationships

  • Jiawei Han — PhD advisor at UIUC; one of the most influential figures in data mining, whose work on knowledge discovery from networks directly shaped Gu’s dissertation and early research agenda.
  • Zaixiang Zheng — Principal collaborator at ByteDance Seed; project lead on DPLM, DPLM-2, SeedProteo, and related protein modeling systems.
  • Xinyou Wang — Key co-author on the DPLM and DPLM-2 papers (ICML 2024, ICLR 2025), central to building the protein language model series.
  • Huizhuo Yuan — Frequent co-author across both UCLA (MARS optimizer, SPPO) and ByteDance work; a key figure in Gu’s LLM and optimization research thread.
  • Dongruo Zhou — Prolific co-author on reinforcement learning theory (multiple ICML and NeurIPS papers) and nonconvex optimization; now an independent researcher.
  • Difan Zou — Co-author on a series of influential papers analyzing SGD implicit regularization and benign overfitting in overparameterized models.
  • Zixiang Chen — Co-author on deep learning theory and the Self-Play Fine-Tuning line of work (SPIN, ICML 2024).

Personal Style

Gu is unusual among ML researchers for sustaining high-output rigorous theoretical work in academia while simultaneously driving applied engineering at the frontier of industry AI—a mode of operation that typically requires trade-offs he has appeared to resist. His research style across both settings is characterized by a preference for minimax-optimal statistical guarantees rather than heuristic performance improvements, even in practically motivated settings. Publicly, he has been a vocal advocate for the view that AI is not merely a tool for research but is actively restructuring the research pipeline itself—his EurekaClaw project and commentary on X reflect a conviction that “system-level discovery,” where humans, models, and tools co-evolve, is already underway. His departure from ByteDance with the terse sign-off “The best model is yet to come. Scaling continues” was widely read as a statement of intention rather than retrospection.

References