A new arXiv paper explores using generative diffusion models to forecast influenza incidence across space and time. The authors argue that existing mechanistic and statistical methods often fail to capture complex epidemic dynamics, and propose a generative approach as an alternative. The work targets public health planning, where better short-term forecasts can inform preparedness decisions.
A revised arXiv paper examines why masked diffusion language models can produce incoherent text blocks: they decode every position in a block in parallel from separate marginal distributions. The authors propose a coupled-noise distillation approach intended to let such models generate a whole block in a single step while keeping the tokens mutually consistent. The work is a research preprint and has not been peer-reviewed.
A new arXiv paper introduces DenMark, a semantic watermarking approach that embeds signals in the meaning of generated text rather than in specific token choices. This design aims to survive paraphrasing and other edits that preserve meaning, which typically break surface-level watermarks. The work focuses on diffusion language models, a class largely unaddressed by existing semantic watermarking methods built for autoregressive models.
A new arXiv paper introduces ReCAST, a method for assigning credit to individual timesteps when fine-tuning diffusion models with reinforcement learning from multiple reward signals. The work separates how much each reward should influence training, based on user preference, from how informative that reward actually is at a given point in the generation process. This distinction is used to drive online diffusion RL more effectively.
A new arXiv paper introduces Abstract-LoRA, a technique that adapts diffusion models for style transfer using only a single reference image. Rather than fine-tuning the whole network, the approach concentrates training on selected U-Net blocks, which the authors present as a way to bypass the data demands of existing multi-image style transfer pipelines. The work is a preprint and has not yet been peer reviewed.
A new arXiv paper introduces LLaDA-UI, a method that adapts diffusion large language models to vision-language agents that operate graphical user interfaces. Diffusion decoding generates tokens in parallel blocks and in arbitrary order, which the authors argue suits latency-sensitive GUI tasks. The work positions interface agents as a testbed for this alternative to standard left-to-right generation.
Researchers propose RAIN, a region-aware inversion network that recovers the initial noise used to embed semantic watermarks in diffusion models. The method aims to avoid the multi-step diffusion inversion that current Gaussian-Shading extraction typically requires, while leaving image quality largely intact. It is described in a cross-listed arXiv paper.
A new arXiv preprint proposes a diffusion-based multiple-shooting method for indirect optimal control, aimed at generating fuel-efficient spacecraft trajectories. The authors argue that prior diffusion-based control work has largely neglected optimality guarantees, which their approach seeks to address. The work sits at the intersection of generative modeling, robotics-style control, and aerospace trajectory design.
A new arXiv paper proposes DSS, a method for removing unwanted concepts from text-to-image diffusion models. The approach steers the model's internal semantic representations dynamically rather than relying on static interventions, aiming to make erasure more robust against adversarial prompts. The work targets safety concerns such as generated NSFW material or copyrighted content.
A new arXiv paper introduces an adaptive sampling method for masked diffusion language models that decides which tokens to commit during each denoising step. The approach targets a known failure mode where individually confident positions become unsafe when decoded in parallel, aiming to preserve accuracy while still reducing the number of inference iterations. The work is a revision of an earlier preprint and has not been peer reviewed.
A new arXiv preprint proposes a restricted initialization scheme for flow-matching diffusion models based on sparse matrix decomposition, aimed at reducing the cost of adapting these models to downstream tasks. The work builds on the observation that fine-tuning flow-matching models is expensive, and that low-rank adaptation combined with timestep-aware choices may help. The abstract suggests the method constrains initialization at a principal timestep to improve training efficiency.
A new arXiv paper surveys how diffusion-based learning methods can be applied to robot learning, framing the problem through ergodic control and controlled diffusion. It is written as a combined review and tutorial, intended to give researchers a structured entry point to the underlying statistical machinery. The work is cross-listed in machine learning and focuses on deriving complex distributions from data for control tasks.
A new arXiv paper introduces Impute-EM, a diffusion-based approach for filling in missing values in datasets that mix numerical, categorical, and binary variables. Most existing diffusion imputation methods convert discrete variables into continuous stand-ins, which the authors argue is a limitation the native mixed-state design avoids. The work targets heterogeneous data mining settings where such mixed variable types commonly appear together.
A new arXiv preprint introduces DiTAR+, a dual-optimization approach for continuous-latent autoregressive diffusion transformer models used in zero-shot speech generation. The method targets the limited decoding stability these models show when producing long utterances or handling complex linguistic input. No results beyond the abstract are described in the report.
A new arXiv paper proposes Temporal Self-Distillation, a training method aimed at discrete diffusion language models that generate several tokens at once. Such models lose quality when too many tokens are decoded in parallel, and the technique is presented as a simple way to reduce that degradation. The approach targets faster inference without the accuracy drop that usually accompanies aggressive parallel decoding.
A new arXiv paper introduces ProtoGuide, a method for steering class-conditional graph generation without baking the class label into the denoiser during training. Instead of embedding the conditioning signal into the model, the approach uses prototypes to guide sampling, which decouples the conditioning mechanism from any particular trained model. This makes it possible to add or change class conditioning on top of existing discrete diffusion generators rather than retraining them.
A new arXiv paper introduces DA-DLM, a method for diffusion language models that explicitly captures relationships between tokens during generation. Existing diffusion models denoise masked text by predicting several tokens at once under an assumption of conditional independence, which the authors say weakens coherence. The proposed approach aims to restore those inter-token dependencies.
A new arXiv paper proposes using diffusion models to synthesize musculoskeletal gait trajectories tailored to individual patient parameters, a step toward better reference signals for wearable assistive devices. The authors frame this as a challenge for lower-limb exoskeletons and rehabilitation robotics, where control depends on trajectories matched to each user. The abstract is preliminary and reports no experimental results yet.
A new arXiv paper argues that the action-generation backbones in Vision-Language-Action models are far larger than the task requires, since robot actions carry much less information than image pixels. The authors introduce a freeze-share-shrink strategy to cut parameters in diffusion and flow-matching policies while preserving performance. The work targets more efficient manipulation models for robotics.
A new arXiv preprint reviews how diffusion-based large language models can be applied to code generation, an area currently dominated by left-to-right autoregressive decoding. The authors examine the limitations of standard autoregressive generation and assess whether diffusion approaches offer advantages for producing source code. The work is a replacement submission (v3) to the cs.AI category.
A new arXiv preprint develops a theoretical account of why diffusion models sometimes reproduce training data verbatim rather than generating novel samples. The authors propose smoothing the score function as a way to reduce this memorization effect and improve generalization. The work is presented as an explanation of the phenomenon rather than a new model release.
A new arXiv preprint introduces a technique called Adaptive Diffusion Freezing that aims to make diffusion models more resistant to membership inference attacks, which try to determine whether a specific sample was part of the training data. The method adapts how parts of the model are frozen during training to limit the privacy leakage that standard diffusion training can expose.
Researchers have introduced a conditional diffusion framework that refines coarse ERA5 reanalysis fields into fine-scale air temperature estimates by incorporating sparse ground-station measurements. The approach is designed to capture terrain and land-surface contrasts that coarse products miss, with the goal of improving local heatwave hazard assessment in areas with limited sensor coverage.
A new arXiv paper proposes an auditing instrument built on causal abstraction to assess fairness in text-to-image diffusion models. The approach aims to avoid the heavy computation normally required when generating many images across different sampling configurations. It targets making comprehensive fairness evaluations more practical for these models.
A new arXiv paper introduces AcFlow, a technique that controls text-to-image diffusion transformers by learning conditional activation flow at inference time. The approach aims to give users adjustable control over style intensity and to suppress unwanted concepts that direct prompting handles poorly. The work targets the limits of prompt-based interfaces for generative image models.
A new arXiv paper reports a controlled experiment on CIFAR-10 using a DDPM U-Net to measure how LoRA rank affects fine-tuning quality and compute cost. The authors tested ranks of 2, 4, 8, 16, and 32 under fixed optimization settings and evaluated results with PyTorch-FID in a reproducible setup. The work aims to give practitioners clearer guidance on choosing a rank that balances output quality against training expense.
A new arXiv paper examines whether diffusion transformers can generate useful cross-modal training data for brain state decoding. The authors note that prior work has mostly fused paired modalities for prediction rather than using their correspondence to augment training data. The approach aims to improve multimodal representation learning.
A new arXiv paper proposes a decoding-control method for diffusion vision-language models that uses intermediate answer trajectories to decide how much reasoning a query actually needs. The approach targets a mismatch in which models apply reasoning effort that does not match the difficulty of the task, aiming to make inference more efficient and better calibrated. The work is presented as a preprint and has not yet been peer reviewed.
A new arXiv paper proposes SAC-Copula, a watermarking scheme designed for diffusion language models, which generate text through iterative parallel unmasking rather than left-to-right decoding. The authors argue that existing sampling-based watermarking techniques, which add independent noise at each position, are poorly suited to this parallel process. Their approach instead uses smooth, correlated Gumbel fields intended to preserve output quality while embedding a detectable signal.