Foundations of Deep Generative Models

Overview

Why This Workshop Now

Deep generative models now power image synthesis, video generation, language modeling, and scientific discovery, but their most important capabilities are still poorly understood.

Diffusion models, flow-based models, and autoregressive models have delivered impressive empirical gains. At the same time, major open questions remain around reliability, interpretability, privacy, and scientific use. This workshop creates a focused venue for theory, empirical analysis, and domain-driven applications to meet.

The program is designed around a practical scientific question: when a model appears capable, is it reproducing training data, capturing a real distributional structure, or performing a stronger form of compositional reasoning that transfers beyond what it has seen?

Memorization

Understand how over-parameterized DGMs retain, reproduce, or expose training data, and what that means for privacy, robustness, and trustworthy deployment.

Generalization

Characterize when generated samples reflect learned structure rather than template matching, and how model size, data complexity, and training dynamics shape that boundary.

Reasoning

Evaluate whether DGMs can support compositional, causal, or structured inference that matters for multi-step generation and scientific workflows.

Topics of Interest

The workshop welcomes work on foundational, empirical, and application-driven questions in deep generative models.

This workshop aims to bring together researchers working on the foundations of diffusion models, flow-based models, autoregressive models, and related generative learning frameworks. We welcome work that sharpens our understanding of what DGMs learn, how they behave under scale, and how they can be evaluated reliably.

The topics below reflect the research directions prioritized for the workshop and are intended to guide the scope of submissions and discussion during the workshop.

01

Memorization and Generalization

Empirical and theoretical studies of memorization, generalization, regime transitions, and the roles of capacity, data complexity, and scaling.

02

Reasoning and Compositionality

Mechanisms for compositional, causal, or structured inference, including in-context learning, chain-of-thought, and multi-step generation.

03

Optimization and Inductive Bias

Learning dynamics, architecture, and implicit regularization in shaping memorization, generalization, and reasoning behavior.

04

Evaluation and Benchmarking

Metrics and diagnostic frameworks for distinguishing memorization from genuine generalization, plus robustness, privacy, and extrapolation benchmarks.

05

Scientific Discovery with DGMs

Applications in scientific machine learning, healthcare, protein design, and molecular discovery where interpretability and reasoning matter as much as raw sample quality.

Call for Papers

Submission logistics and formatting requirements are still being finalized.

Important Dates

Paper Submission Deadline: April 31, 2026
Notification of Acceptance: May 15, 2026
Camera-Ready Deadline: June 5, 2026
Workshop Date: TBD (Either July 10 or 11)

Submission Instructions

The official submission portal will be posted here once it is available.
Contributed papers are expected to align with the workshop scope described in Topics of Interest.
Accepted submissions are expected to appear as posters, with a subset selected for short oral presentations.

Additional logistics, including review process details and camera-ready instructions, will be added once confirmed.

Formatting Instructions

Paper format, page limits, and anonymization requirements are to be announced.
Templates and camera-ready instructions will be posted here once available.
Please check back for finalized guidelines before submission opens.

Schedule

Workshop Schedule

This workshop combines invited talks, contributed oral presentations, poster sessions, a panel discussion, and closing awards.

Format at a glance

One-day in-person workshop within the ICML 2026 workshop program.
Six invited talks spanning theory, reasoning, and scientific applications.
Poster sessions and short oral slots reserved for contributed work.
Panel discussion focused on open problems and future research directions.

Morning

8:20-8:30Opening remarks
8:30-9:00Invited talk 1
9:00-9:30Invited talk 2
9:30-10:30Poster session and break
10:30-11:00Invited talk 3
11:00-11:30Invited talk 4
11:30-11:45Oral presentation 1
11:45-12:00Oral presentation 2

Afternoon

12:00-1:30Lunch
1:30-2:00Invited talk 5
2:00-2:30Invited talk 6
2:30-3:30Poster session and break
3:30-3:45Oral presentation 3
3:45-4:00Oral presentation 4
4:00-4:50Panel discussion
4:50-5:00Awards and closing

Speakers

Confirmed Invited Speakers (Alphabetical Order)

All seven invited speakers listed here are confirmed. The speaker slate spans theory, reasoning, optimization, and scientific applications of deep generative models across multiple continents and career stages.

Kenji Fukumizu

Professor, The Institute of Statistical Mathematics

Representative topics: proximal flow matching and flow matching generalization.

Surya Ganguli

Associate Professor, Stanford University

Representative topics: mathematical reasoning and creativity and inductive biases.

Ge Liu

Assistant Professor, University of Illinois Urbana-Champaign

Representative topics: protein design and flow matching on structured spaces.

Yi Ma

Chair Professor, University of Hong Kong

Representative topics: memorization in DGMs and post-training generalization.

Cengiz Pehlevan

Associate Professor, Harvard University

Representative topics: in-context learning theory and spectral bias and learning dynamics.

Taiji Suzuki

Professor, University of Tokyo

Representative topics: flow matching and chain-of-thought reasoning.

Mengdi Wang

Professor, Princeton University

Representative topics: score approximation and generative models for biology.

Organizers

Organizing Team (Alphabetical Order)

The team combines expertise across diffusion models, deep learning theory, optimization, sampling, and applications.

Beatrice Achilli

Postdoctoral Researcher, Bocconi University

Her research lies at the intersection of statistical physics and machine learning, with a particular focus on diffusion models.

Valentin De Bortoli

Research Scientist, Google DeepMind

His research lies at the intersection of stochastic control, optimal transport, probability, statistics, and generative modeling. He has organized workshops across NeurIPS, ICML, and scientific computing communities.

Wei Huang

Research Scientist, RIKEN Center for Advanced Intelligence Project

His research focuses on the theoretical foundations of language models and deep learning. He has co-led the DeLTa workshops and co-organized related events in deep learning theory.

Qing Qu

Assistant Professor, University of Michigan

His research centers on the foundations of data science, optimization, deep representation learning, and diffusion models. He has led major community efforts including CPAL, DeepMath, and multiple tutorials and workshops on diffusion models.

Molei Tao

Professor, Georgia Institute of Technology

Research spans diffusion models, sampling, deep learning theory, optimization, and AI for science. He has organized workshops at NeurIPS and numerous seminars and minisymposia across applied mathematics and machine learning venues.

John J. Vastola

Postdoctoral Researcher, Harvard University

His research examines generalization in generative models, the biological basis of memory, and principles of efficient learning, with publications in venues including Nature Communications, Cell Systems, NeurIPS, and ICLR.

Peng Wang

Assistant Professor, University of Macau

His research focuses on theoretical machine learning, signal processing, and convergence theory for structured non-convex optimization. He has also served in recent area chair roles at ICLR and CPAL.

Renyuan Xu

Assistant Professor, Stanford University

Her research focuses on the foundations of generative AI, high-stakes decision systems, and machine learning in finance. She previously organized the Generative AI in Finance workshop at NeurIPS 2025.

Student Organizers

Student Organizers (Alphabetical Order)

Awards

The workshop schedule includes an awards-and-closing slot at the end of the day. Additional recognition details for outstanding submissions and presentations will be shared here.

TBD

Contact

Workshop Information

For questions about participation, contributions, or logistics, contact the workshop leads below.

Primary Contacts

FDGM Workshop icml.workshop.dgm@gmail.com
Peng Wang pengw@um.edu.mo
Qing Qu qingqu@umich.edu

Planning Notes

ICML 2026 will be held in Seoul, South Korea, from July 6 to July 11, 2026.
Workshops are scheduled for July 10 and July 11, 2026. The exact day for this workshop is to be announced.
The workshop anticipates strong participation from both theory and applied DGM communities.