ICML 2026 Workshop: Foundations of Deep Generative Models — Theoretical Exploration of Memory, Generalization, and Reasoning

This article introduces the ICML 2026 Workshop on Foundations of Deep Generative Models (FDGM), focusing on theoretical research progress of deep generative models in three core issues: memory, generalization, and reasoning, and discusses the theoretical foundations and challenges in the era of large language models. The workshop is maintained by fdgm-workshop, hosted on GitHub, and published on June 16, 2026.

Deep generative models have become a core pillar of AI, but theoretical understanding lags behind technological development. As a top-tier conference, ICML's workshops provide a platform for exchanging cutting-edge directions. The FDGM workshop focuses on the theoretical foundations of deep generative models, especially the three core interrelated issues of memory, generalization, and reasoning.

Memory is a controversial topic in generative models, involving privacy risks (leaking sensitive information), copyright disputes (defining similarity between generated content and training data), and capability evaluation (whether memory equals understanding). The workshop discusses progress in applications of differential privacy, defense against membership inference attacks, and information-theoretic quantification of memory capabilities.

Generalization of generative models requires evaluating the difference between the generated distribution and the real distribution (e.g., Wasserstein distance, MMD). Current hot topics include sample complexity (amount of training data needed for high-quality samples), mode coverage (avoiding mode collapse), and out-of-distribution generalization (robustness).

Reasoning capabilities include causal reasoning (understanding variable relationships), compositional generalization (combining known concepts), and multi-step planning (decision-making for complex tasks). The emergent reasoning capabilities of large language models have sparked discussions: do they stem from pattern matching or abstract reasoning?

1. Model safety and alignment: Understanding memory and generalization mechanisms can help design safety protections; 2. Training data strategy: Guiding data deduplication and quality screening; 3. Model architecture: Inspiring the design of next-generation generative models with stronger cognitive capabilities.

Theoretical results have guiding significance for industry: 1. Large model training: Optimizing data ratio to improve efficiency; 2. Content moderation: Generalization theory provides compliance evaluation tools; 3. Product planning: Reasoning research points out the evolution direction of AI assistants.

The FDGM workshop brings together researchers from multiple fields, and interdisciplinary exchanges promote theoretical breakthroughs. Theory is the ballast stone of technology; practitioners' attention to theoretical results can avoid blind expansion and help generative models develop steadily and far.