arXiv:2610. 12865v1 Announce Type: new Abstract: Existing convergence guarantees for non-log-concave sampling with stochastic gradient Langevin Monte Carlo (SG-LMC) typically assume globally Lipschitz gradients and uniformly bounded noise variance.
arXiv:2610. 10190v1 Announce Type: cross Abstract: We analyze the first-order deterministic diffusion sampler of Karras et al.
arXiv:2609.40193v1 Announce Type: new Abstract: We establish near-linear accuracy bounds for the classical Moreau--Yosida unadjusted Langevin algorithm (MYULA). The target is $\pi\propto e^{-f-g}$, w...
arXiv:2506. 13058v2 Announce Type: replace-cross Abstract: Diffusion probabilistic models (DPMs) have demonstrated remarkable success in visual generation.
arXiv:2505.20817v3 Announce Type: replace-cross Abstract: Gradient clipping is widely used in language-model training to control heavy-tailed gradient noise and can improve convergence guarantees ove...
arXiv:2606. 15359v1 Announce Type: new Abstract: Diffusion models have emerged as powerful tools for planning and control by learning multimodal distributions over actions and trajectories.
The paper studies how to allocate a fixed computational budget across the denoising steps of diffusion models to improve sample quality at deployment. It shows that the expected benefit of evaluating multiple candidates at a step can be decomposed into a step‑specific sensitivity and a universal sample‑size factor, and that the optimal allocation follows a water‑filling structure. Experiments demonstrate that this allocation achieves the same quality as a uniform strategy while reducing function evaluations by 20–50%.
arXiv:2607. 12360v1 Announce Type: new Abstract: The cooldown phase of a warmup-stable-decay (WSD) learning-rate schedule, now a default in large-model pretraining, lowers the final training loss in some settings and does nothing in others.
arXiv:2606. 00293v1 Announce Type: new Abstract: Tuning algorithms such as stochastic gradient descent (SGD) and stochastic gradient Langevin dynamics (SGLD) for approximate sampling and uncertainty quantification remains challenging, particularly in the practically relevant settings when the batch size is large or the model is misspecified.
arXiv:2607. 26285v1 Announce Type: cross Abstract: Two central challenges in diffusion-based sampling are the theoretical one of understanding their remarkable effectiveness even in high-dimensional settings, and the practical one of designing algorithms with certified performance guarantees.
The paper investigates whether momentum-based acceleration can improve Gaussian variational inference (GVI) for strongly log-concave, log-smooth targets. With exact Gaussian expectations, GVI achieves the classic accelerated √κ convergence rate, but when only pointwise derivatives are available, the authors construct targets that force any randomized algorithm to use κ^{1-o(1)} queries, showing curvature alone does not guarantee acceleration. They identify additional structure—specifically a Hessian-Lipschitz condition with small Hessian variation—under which variance-reduced stochastic methods can achieve accelerated guarantees, and demonstrate empirical speedups on real datasets.
The paper proves that for discrete diffusion models using uniform or remasking forward processes, an adaptive sampler based on a leave‑one‑out denoiser can achieve sampling error proportional to the score‑estimation error plus a small tolerance. The required number of discretization steps scales with the dual total correlation of the target distribution, not directly with the ambient dimension. This result shows that sampling complexity is governed by the intrinsic dependence structure of the distribution, and the authors provide an information‑theoretic analysis linking discretization error to mutual information between coordinates.