flow-matching

University of Missouri researchers published a comprehensive review of flow matching, an emerging generative AI approach for biology that models how living systems change over time, providing a roadmap to accelerate drug discovery, precision medicine, and the development of AI-powered virtual cells.

提出了异步流映射语言模型（aFMLM），通过token级时间调度将预测顺序编码进单步生成的隐藏计算中，在Sudoku和GSM8K代码生成任务上刷新了一步生成的最先进性能。

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 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.

WorldBridge introduces a World Action Model with separate video and action experts connected via structured attention and trained with joint flow matching, achieving state-of-the-art task generalization on robot manipulation benchmarks by combining broad video coverage with sparse action supervision.

AuroraAI-Research 发布了 AuroraIMG2-50M-Preview，一个基于 5040 万参数流匹配扩散 Transformer 的 256×256 文生图模型，搭配 1260 万参数文本编码器，全部从零训练仅需单卡 RTX PRO 6000 上 10 小时完成。

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.

论文提出 CD-FFM，一种基于似然函数流匹配（likelihood functional flow matching）的生成式气候降尺度模型，可将再分析数据降尺度至1公里分辨率，在芬诺斯坎迪亚日气温数据上优于回归、神经算子和扩散模型，且采样速度极快并能给出精确似然。

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.

GLOW-FM is a one-step latent flow-matching framework for efficient generative low-light image enhancement, performing conditional transport in a frozen Tiny-VAE latent space with Terminal Velocity Matching to achieve state-of-the-art cross-dataset generalization using a single function evaluation and 34 GFLOPs.

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.

ArcFlow 提出一种部署预算感知的训练算法，使单个流策略（flow policy）能在多个求解器步数预算下保持高性能，尤其在一步推理场景中优于五个基线方法。

StyleStream 2.0 是一个快速且可控的流式语音风格转换框架，通过 few-step MeanFlow 模型和自回归反馈微调将端到端延迟降至 520ms，并借助统一的文本条件流匹配模型支持文本驱动的语音设计与指令式语音编辑。

StreamingSwap is a real-time flow-matching framework for video face swapping that uses a Joint Rec-ID loss and context-efficient generation (Face-RoI + short-context rollout) to achieve 20 FPS on an RTX 4090 over videos up to 10 minutes, along with the HumanFaceVid 1,000-hour cinematic dataset.

Fréchet Forcing 提出了一种无需教师模型和长视频的微调方法，利用自生成轨迹与真实数据统计量（生成锚定、上下文保真、现实匹配三项约束）训练流式视频世界模型，使其在超过 4000 帧的长时程推理中保持稳定并正确跟随动作指令。

DeOverlay 提出一种基于源锚定流匹配与可见性归一化时序聚合的视频水印/字幕去除方法，并发布 DVTE-Bench 基准（1,631 对视频、八类文字覆盖），在五项重建与感知指标上显著优于四个基线。

VLA-Rolling is a lightweight inference framework that reuses discarded action tails from flow-matching VLA models with adaptive denoising budgets, achieving up to 1.48x speedup and up to 15.08 percentage-point success rate gains across 16 model-benchmark pairs on LIBERO, MetaWorld, LIBERO-Plus, and RoboTwin 2.0.

AffordanceWAM 提出一种感知可供性（affordance）的生成式世界-动作模型，利用人类视频中的可供性表征联合预测未来场景与机器人动作，无需人类动作标注即可实现人类到机器人的迁移，并在 RoboCasa、CALVIN 等基准上取得一致提升。

Kyutai Labs introduces REKI, a 2.4B vision-language-action model pretrained exclusively on public data that sets a new Pareto frontier between performance and model size, using a simple design with register-token memory in the action expert. Full code, data curation pipelines, and all pretrained/finetuned weights are released.