Tag
The paper presents PIFT, a framework that transfers peptide–target interaction patterns to small-molecule generation via joint sampling-time guidance of masked discrete diffusion and flow matching, targeting challenging protein-protein interaction interfaces with strong docking results.
The paper introduces a divergence-kernel method for jointly training neural SDE drift and diffusion fields with losses that depend directly on the terminal density, achieving fewer off-mode samples than score-based generators on Gaussian-mixture benchmarks with time-independent fields.
The paper proposes PoSE, a Bayesian distribution system state estimation method that uses a flow-matching-trained neural network to learn only the distribution of substation injections, while power flow equations and measurements enter at sampling time via a noise-aware Kalman-type correction, achieving better CRPS and calibrated uncertainty than Gaussian priors and supervised estimators on IEEE 33-bus and Iowa 240-bus feeders.
该论文提出扩散Transformer的训练加速部分源于AdaLN-Zero残差门控α的重组，并引入归一化参与度比(nPR)和一种无教师的径向目标，在ImageNet 256x256上以SiT-B/2改进了FID。
The paper introduces Variational Atlas Autoencoders, a mixture-of-charts framework that learns multiple local latent representations over a data manifold, using latent-evidence routing to treat chart assignment as self-supervised posterior inference, achieving better-structured latent spaces and higher sample quality on datasets like CelebA.
This paper benchmarks six generative deep learning models for nowcasting aerosol particle size distributions from PM2.5 and meteorological data, finding that physics-guided flow matching performs best, with a reduced general dynamic equation constraint lowering CRPS by 5.5–11.5% at the training site.
A new paper introduces LPDS, a generative inverse modeling framework that maps desired fusion outcomes and target configurations to optimized laser pulse shapes for inertial confinement fusion, using multi-objective physics-constrained training and control mechanisms like constraint-conditioning and inpainting.
The paper identifies a compatibility failure mode in likelihood-based Riemannian flow matching on manifolds, gives a general condition for compatibility, and proposes a new exponential-family-over-endpoints objective that is both compatible and supported only on feasible velocities, achieving improved NLL and sample quality on tori, sphere benchmarks, and human pose datasets.
该论文提出 Transformer Energy-based Model (TEM)，通过在潜空间中构建可扩展的 Transformer 能量函数并多层能量聚合，实现了无需 classifier-free guidance 的高质量图像生成（ImageNet 256×256 FID-50K 达 3.82），验证了 EBM 的组合生成、零样本修复与分布外检测能力。
The paper proposes Doob Policy Optimization (DoobPO), a likelihood-free RL fine-tuning framework for diffusion and flow-based generative policies that uses Doob's h-transform to learn a ratio model and distill the tilted target policy via conditional drift matching, achieving strong results on MuJoCo continuous-control benchmarks while supporting PPO and GRPO-style updates.
This paper studies predictive flow matching under distribution drift, proposing DyProd-FM which refits on dyadic historical prefixes and uses a sleeping Prod master to achieve minimax-optimal dynamic regret under both relative velocity risk variation and target-law variation, without requiring convexity or curvature assumptions.
This paper investigates why the standard Gaussian prior N(0,I) on VAE latent spaces can fail under RL fine-tuning of latent diffusion models, showing that KL regularization compresses multimodal data into overlapping latent regions, and demonstrating a 71× improvement in collapse threshold using a Gaussian-mixture prior via a geometric mediator, with Lean 4-verified theory.
ORBIT 提出了一种统一的深度时间序列生成框架，通过频率层次化 VQ-VAE 和频率带自回归 Transformer 实现从低频到高频的下一频带预测，从而先构建全局时间结构再逐步补充细节，在多个真实基准上达到最先进的性能与更高的推理效率。
RealMe 是一个基于高分辨率视频扩散模型的统一框架，能从稀疏、低分辨率的多视角采集生成 2K 分辨率且 4D 一致的沉浸式体积视频，通过布局感知注意力和频域潜空间监督，在 DNA-Rendering 基准和自建 RealVV 数据集上超越现有方法。
提出 acquisition function-guided generation (AFGG) 方法，通过将采集函数倾斜的分布嵌入扩散桥，在离散空间上结合预训练生成模型与数据空间代理模型进行贝叶斯优化；在分子设计基准上，基于流的变体在多数场景中超越现有最优方法。
The paper introduces Kinematic Interpolants (KI), a framework extending stochastic interpolants to infer nonlinear dynamics from phase-space snapshots containing both position and velocity measurements, achieving better reconstruction with far less training time across scientific datasets.
提出了GOLD框架，通过GeoVAE将RNA离散序列与连续侧链角度投影到统一潜在流形上进行协同扩散，并引入训练无关的梯度引导采样机制，实现结构感知的RNA逆折叠，达到最先进性能。
The paper introduces LOB-Impact, an open-source benchmark showing that neural generative limit order book models trained on next-token prediction exhibit emergent market impact when metaorders are injected, with impact exponents between 0.63 and 1.21, though relaxation decay remains unobserved likely due to short context windows.
Halcyon demonstrates that strong text-to-image generation can be trained from scratch on image-text pairs without large-scale language pretraining, outperforming existing models with far less compute. Controlled experiments show text-only pretraining does not improve T2I generation.
Holo4D is a video reshooting framework that uses reconstructed 3D trajectory scaffolds to guide a pretrained video diffusion model for novel-view video synthesis, preserving scene geometry and motion while generatively completing unseen regions without retraining the generator.