The paper investigates when language diffusion models, specifically Uniform-based Discrete Diffusion Models (UDDMs), shift from memorizing training data to generalizing to new data. It shows that UDDMs act as associative memories, forming basins of attraction around stored examples without requiring an explicit energy function. By measuring token recovery and conditional entropy, the authors identify a sharp transition governed by training set size, where memorization (vanishing entropy) gives way to generalization (finite entropy).
arXiv:2610.19042v1 Announce Type: new Abstract: Understanding the mechanism behind why diffusion models memorize their training data remains an open question in generative modeling. In this paper, we...
Understanding the mechanism behind why diffusion models memorize their training data remains an open question in generative modeling. In this paper, we examine this question from two perspectives: (a)...
The paper demonstrates that diffusion models encode memorization in the geometry of their learned energy landscape before it becomes visible in generated samples. By analyzing score divergence and basin volume, the authors show that localized basins form around training images early in training, and that cyclic denoising can recover these images even when one‑shot generation shows no copies. Experiments on Gaussian mixtures, CelebA, CIFAR‑10, and Stable Diffusion v1.4 confirm that basin geometry reveals memorization across various settings.
The paper introduces an information-theoretic measure called model information to quantify how much a diffusion model’s outputs depend on its specific training dataset. By analyzing how this measure scales with model size and dataset size, the authors identify two distinct regimes: a nonlinear regime where memorization transitions to imperfect generalization, and a linear regime where generalization improves and sample quality rises. They propose a scaling law that captures these regimes and demonstrate that it can predict memorization and sample quality beyond the fitted data range.
Diffusion language models (dLLMs) have emerged as a promising alternative to autoregressive (AR) language models, offering flexible token-update orders and parallel decoding. Recent dLLMs are often in...
arXiv:2310. 05264v5 Announce Type: replace Abstract: In this work, we investigate an intriguing and prevalent phenomenon of diffusion models which we term as "consistent model reproducibility": given the same starting noise input and a deterministic sampler, different diffusion models often yield remarkably similar outputs.
arXiv:2609.37974v1 Announce Type: cross Abstract: Masked diffusion models (MDMs) generate text by unmasking several tokens per step, but they are trained and sampled under different conditions. The m...
arXiv:2610.08108v1 Announce Type: new Abstract: Diffusion language models (dLLMs) have emerged as a promising alternative to autoregressive (AR) language models, offering flexible token-update orders...
How diffusion models circumvent the curse of dimensionality to learn complex distributions over high dimensional spaces from a finite training set, instead of memorizing it, remains a fundamental mystery. To address this, we introduce analytically tractable Bayesian information restricted diffusion (BIRD) models, in which each pixel observes restricted information about noisy data.
arXiv:2607. 08041v1 Announce Type: new Abstract: How diffusion models circumvent the curse of dimensionality to learn complex distributions over high dimensional spaces from a finite training set, instead of memorizing it, remains a fundamental mystery.
The paper investigates two key design levers for Continuous Diffusion Language Models (CDLMs): self‑conditioning and embedding geometry. It shows that self‑conditioning mitigates a train/inference mismatch and can be implemented efficiently via a two‑level interpolant called DoubleCat, which significantly boosts TinyGSM accuracy. By adding discrete reveals to DoubleCat, the authors create SemiCat, achieving discrete‑diffusion‑level performance and outperforming discrete diffusion on several likelihood‑based benchmarks at 1.7B parameters. The study also finds that cross‑entropy training drives embeddings toward a geometry with decoding error rates comparable to one‑hot representations, and briefly explores additional options such as hybrid corruption processes and asynchronous noise schedules.
arXiv:2606. 24543v1 Announce Type: new Abstract: Large Language Models (LLMs) are traditionally viewed as autoregressive generators.