NeurIPS 2026 Papers

Many training-free sparse attention methods are effective for accelerating diffusion models. Recently, several works suggest that making sparse attention trainable can further increase sparsity while preserving generation quality. We study three key questions: (1) when do the two common masking rules, i.e., Top-k and Top-p, fail, and how can we avoid these failures? (2) why can trainable sparse attention reach higher sparsity than training-free methods? (3) what are the limitations of fine-tuning sparse attention using the diffusion loss, and how can we address them? Based on this analysis, we propose SpargeAttention2, a trainable sparse attention method that achieves high sparsity without degrading generation quality. SpargeAttention2 includes (i) a hybrid masking rule that combines Top-k and Top-p for more robust masking at high sparsity, (ii) an efficient trainable sparse attention implementation, and (iii) a distillation-inspired fine-tuning objective to better preserve generation quality during fine-tuning using sparse attention. Experiments on video diffusion models show that SpargeAttention2 reaches 95% attention sparsity and a $16.2\times$ attention speedup while maintaining generation quality, consistently outperforming prior sparse attention methods.

Retrieval-augmented diffusion models (RAG-DMs) have significantly advanced image synthesis by incorporating external knowledge bases. However, their security vulnerabilities against malicious external data remain largely underexplored. Existing data poisoning attacks against diffusion models primarily corrupt internal parameters. They fail against RAG-DMs because clean retrieved images easily override these compromised parameters. Furthermore, recent attack tailored for RAG-DMs requires training the retriever, resulting in high computational costs and overfitting to specific retrievers. To bridge this gap, we propose PoisonedRDM, a novel and train-free data poisoning attack tailored to execute concept hijacking against RAG-DMs. Specifically, PoisonedRDM poisons the knowledge base with a minimal number of optimized images and optimizes adversarial perturbations through a joint optimization strategy, steering unknown retrievers toward the poisoned images via surrogate ensemble alignment and aligning the generated outputs with a malicious target concept, while adhering to a strict perturbation budget to maintain high visual stealthiness. Experiments show that PoisonedRDM effectively attacks RAG-DMs, achieving high success rates and outperforming state-of-the-art baselines.

We introduce Track4D, a method that formulates dense 3D tracking as conditional video generation. Repurposing large-scale pretrained video generators for 3D tracking is non-trivial: 3D point tracking in world space produces signals far from natural video, making it difficult for the video diffusion model (VDM) to learn. We systematically study two representations for encoding 3D tracking in the VDM latent space: a Residual 3D Tracking Video (RTV) that directly encodes metric 3D offsets, and a Normalized Coordinate Map Video (NCMV) that implicitly encodes the underlying geometric correspondences as canonical 2D coordinate maps. We further condition the VDM on point maps from an off-the-shelf 3D reconstructor to provide an explicit geometric scaffold for tracking. Trained exclusively on limited synthetic data with LoRA adaptation, Track4D achieves competitive zero-shot tracking performance across five benchmarks, demonstrating the potential of leveraging video generation priors for dense 3D tracking.

A key strength of diffusion models lies in their flexibility, since their outputs can be controlled at sampling time through guidance. However, beyond simple cases such as conditional sampling, the target distribution is often left implicit, defined only through a sampling rule or a heuristic energy function. To address this, we propose Jeffrey guidance, a principled framework that extends diffusion-model control to applications beyond what standard guidance can express. It leverages Jeffrey’s rule of conditioning to update marginal distributions towards a prescribed target, preserving the conditional structure and minimally perturbing the joint distribution. We first demonstrate Jeffrey guidance by targeting a prescribed embedding distribution. With Inception embeddings as the target, this leads to substantial reductions in FID on both CIFAR-10 and FFHQ. We further apply Jeffrey guidance to fairness on CelebA-HQ, updating an unconditional diffusion model to enforce independence between attributes.

Diffusion language models (DLMs) promise parallel, order-agnostic generation, but on standard benchmarks they have historically lagged behind autoregressive models in sample quality and diversity. Recent continuous flow-matching approaches over one-hot and embedding spaces have narrowed this gap, suggesting continuous state spaces are highly effective for language. In this work, we further close the autoregressive gap by modeling text as a continuous diffusion process over fixed-width bitstreams. Our approach represents semantic tokens as analog bit sequences and utilizes a matched-filter residual parameterization to isolate contextual learning from analytic independent-bit posteriors. Crucially, our empirical comparisons reveal that while deterministic continuous DLMs are competitive, they can be overly contractive, undershooting real-data entropy. We demonstrate that the remaining quality-diversity gap is bridged by a stochastic sampler that applies Langevin-type corrections gated by the entropy-rate profile, automatically concentrating stochasticity in high-information regions while remaining nearly deterministic elsewhere. On the One Billion Word Benchmark (LM1B), our 130M-parameter bitstream model reaches a generative perplexity (GenPPL) of 59.76 at matched real-data entropy (4.31) using 256 neural function evaluations (NFEs), decisively outperforming prior DLM baselines and reaching the autoregressive reference. On OpenWebText (OWT), our stochastic sampler establishes a new continuous-DLM Pareto frontier, achieving GenPPL = 27.06 at an entropy of 5.26 using 4x fewer steps than previous 1024-NFE baselines. As an additional architectural benefit, bitstream diffusion removes the O(V) vocabulary scaling bottleneck shared by standard DLMs. By predicting O(log V) bitwise logits via semantic bit-patching, our model yields a reduced memory footprint and higher throughput, demonstrating a scalable paradigm for language generation as vocabulary sizes grow.

Large Language Models (LLMs) have demonstrated remarkable capabilities in orchestrating tools for reasoning tasks. However, existing methods rely on a step-wise paradigm that lacks a global perspective, which causes error accumulation over long horizons and restricts generalization to unseen tools. To overcome these limitations, we propose Tools as Continuous Flow for Evolving Agentic Reasoning (FlowAgent), which reconceptualizes tool chaining as continuous trajectory generation within a semantic space. To systematically evaluate this paradigm, we introduce the first plan-level closed-loop benchmark dedicated to plan-level agentic reasoning in dynamic real-world environments. Specifically, the proposed FlowAgent leverages conditional flow matching to generate continuous latent trajectories, providing a global planning perspective to ensure coherent and robust tool execution. Theoretically, we establish formal bounds on utility convergence and prove that our continuous formulation fundamentally guarantees robust generalization and error attenuation. Empirical evaluations show that FlowAgent achieves superior robustness and adaptability in long-horizon reasoning tasks.

AI-generated image detection has become increasingly important as generative models produce highly realistic visual content. Existing detectors mainly rely on visual artifacts or visual features extracted from pretrained multimodal models, while largely overlooking the structural differences between real and generated images in the joint image--text space. In this paper, we reveal a systematic Cross-modal Alignment Shift (CAS): generated images tend to exhibit stronger and more concentrated alignment with text than semantically matched real images. We verify this phenomenon through retrieval preference, image--text residual structure, spectral concentration, and matching entropy. Motivated by this finding, we propose a CAS-based detection framework(CASNet) that learns an amplified multimodal model with a text-guided alignment-shift objective. The amplification-induced feature increment is extracted as an explicit cross-modal cue and coupled with amplified visual representations, enabling the detector to exploit both alignment-level structural evidence and visual discriminative cues. Extensive experiments across diffusion models, GANs, autoregressive models, and commercial generators demonstrate consistent improvements over state-of-the-art detectors. In particular, on recent commercial generators, our method improves ACC and AP by 5.48\% and 6.13\%, respectively.

We present an inference-time approach for enhancing the visual diversity of image sets generated by text-to-image (T2I) models, while preserving, or even improving, the average image quality. Our method leverages the intrinsic geometry of hidden latent representations in pretrained models to diversify a batch of images generated under the same prompt. First, we introduce a token repulsion mechanism to encourage diversity by increasing the angle between the tokens of hidden features across different samples. This is achieved via a negative Riemannian gradient step defined over a distribution on the unit hypersphere. Second, to preserve the quality and prompt adherence of the output images, we encourage the updates to remain within the high-density region learned by the generative model. Third, we extend this constraint to the output of each intermediate layer, enabling more fine-grained iterative guidance. We evaluate our method on three representative large-scale pretrained T2I models: Flux1.dev (flow matching), SDXL (diffusion), and Flux2.Klein (few-step distilled flow matching). Experimental results demonstrate substantial improvements in visual diversity under identical prompts, while maintaining averaged individual sample fidelity. Our approach is simple, effective, and broadly applicable. It requires no external models to guide the generation process and introduces only modest additional computation overhead during the early stages of sampling.

Probabilistic long-term time series forecasting requires estimating a distribution over future trajectories under practical deployment constraints, where inference latency and stability can be as critical as forecast quality. Existing diffusion-based forecasters provide expressive uncertainty modeling, yet their iterative denoising incurs substantial horizon insensitive runtime overhead, which limits real time long-term use. In contrast, one-shot generative models such as VAEs and flows are computationally efficient but often struggle to capture complex multi-modal structure and heavy tailed uncertainty. We propose \textbf{ToLD}, an efficient time‑series forecasting framework built upon \textbf{to}kenized truncated \textbf{l}atent \textbf{d}iffusion to enable fast and reliable long term generation. ToLD adopts a tokenized conditional latent modeling and performs compact token-wise refinement via a few-step truncated latent diffusion process, where the diffusion noise is injected in a context-dependent manner to model non-stationary uncertainty. To reduce train test mismatch in multi-sample forecasting, we further introduce a score distillation scheme that learns a target-free scoring function for test-time candidate ranking. Experiments on eight real-world datasets show that ToLD improves both point accuracy and probabilistic quality over state-of-the-art, achieving up to \textbf{4.3\%} MSE improvement and \textbf{19.3\%} CRPS reduction on long-term tasks with minimal additional inference overhead.

Diffusion models, particularly Diffusion Transformers, achieve strong image and video generation quality but remain expensive at inference time due to repeated denoiser evaluations. Feature caching offers a deployment-friendly acceleration strategy by reusing or predicting intermediate representations without retraining the generator or modifying the sampler. However, existing cache-then-forecast methods often rely on Taylor-style finite-difference extrapolation, which can become unstable over longer cache intervals, and typically predict local module outputs whose errors may accumulate through subsequent denoiser blocks. We propose \textbf{BeziCast}, a training-free diffusion acceleration framework that forecasts output-proximal denoising representations using low-order B\'ezier trajectories. BeziCast estimates B\'ezier control points via Tikhonov-stabilized fitting, providing a smooth, capacity-controlled temporal parameterization that avoids explicit high-order derivative estimation and decouples trajectory capacity from cache-query density. To moderate aggressive extrapolation, we further derive equivalent forecasting weights and introduce a convex-hull-inspired guardrail that detects high-risk predictions and softly pulls them toward a conservative simplex-projected estimate. Extensive evaluations across advanced image and video diffusion models demonstrate the effectiveness of BeziCast. In particular, BeziCast accelerates FLUX.1 by up to \(4.79\times\) and HunyuanVideo by up to \(4.11\times\), while preserving substantially better generation quality than competing acceleration baselines.

Over the past decade, a growing body of work has established that implicit biases in learning dynamics fundamentally shape the solutions found by neural networks, governing their learned features and generalisation performance. However, despite these strides in deep learning theory, far less is known about representation learning in generative models based on dynamical systems, from normalising flows to diffusion. Importantly, continuous-time generative models can be prone to mode collapse, yet the learning dynamics underpinning such biases remain elusive. To address this, we use tools from dynamical systems and operator theory to show that bounds on the rank of the gradient of the model weights can explain how collapse occurs in generative models. We show that, in both variational inference and flow matching tasks, optimisation induces low-rank biases that encourage parsimonious representation learning but may also cause the learned distribution to collapse. Together these findings point to a trade-off between promoting feature learning while avoiding both memorisation and collapse in dynamical systems-based generative models.

A central obstacle in nonlinear Bayesian filtering is representing the belief distribution. Moment-based filters address this by propagating polynomial moments and reconstructing a density from them. Recent work completes the predict-update loop via the maximum-entropy (MaxEnt) principle, but each step requires the partition function and its gradient, both $n$-dimensional integrals whose cost scales exponentially, restricting the demonstrated MaxEnt moment filtering to $n \le 4$. We avoid the partition function entirely by combining score matching with Stein's identity. In our setting, score matching reduces the density fit to a single linear solve whose coefficients are assembled directly from the propagated moments. The same parameters then drive Stein's identity to close the moment hierarchy during prediction and to recover posterior moments after each Bayesian update, keeping the full predict-update loop free of partition function evaluation. The resulting $\textbf{Score Kalman Filter}$ (SKF) reduces to the classical information-form Kalman filter as a special case and performs every step through linear algebra. On nonlinear coupled-oscillator networks, the SKF runs through $n=20$ and reports lower RMSE than the EKF, UKF, EnKF, and particle-filter baselines on the tested synthetic benchmarks.

Score- and flow-matching models often rely on preference-based reinforcement learning for two purposes: aligning with subjective preferences and, surprisingly, recovering properties---such as visual realism and coherent object structure---that matching-based training is intended to learn from the data itself. We argue that this reflects a structural mismatch. Matching losses measure $\ell_2$ regression error on the velocity or score field under training-time marginals, a proxy poorly aligned with the visual and semantic properties that determine sample quality at inference. Given a reward aligned with these properties, RL sidesteps the mismatch by evaluating the model on its own samples and following the reward landscape directly. The challenge is to obtain such a reward without relying on human references, which are expensive and conflate data realism with annotator inclinations. We propose Discriminator-Guided RL (DRL). DRL trains a discriminator to separate data from base-model samples in a pretrained representation space and uses its logit as the reward in KL-regularized RL. The pretrained space restricts the discriminator to perceptually meaningful directions, and the logit estimates the log-likelihood ratio between data and model, which is the optimal reward for targeting the data distribution. Across SiT, JiT, REPA, and RAE, DRL reduces guidance-free FID (e.g., $9.38 \to 2.62$ on SiT) and semantic-space FD (e.g., $88.2 \to 19.3$ on DINOv3 for SiT), with consistent gains across all backbones, and improves human-preference rewards without training on them. It also yields a better Pareto frontier between preference reward and image fidelity under subsequent preference-based post-training, increasing alignment while reducing low-level artifacts such as oversaturation and excessive brightness.

Recent text-to-image diffusion models (T2I DMs) depend on vast, uncurated datasets that often include copyrighted artworks and personal images, risking the generation of unwanted content. While concept erasure methods have emerged to suppress such outputs, they predominantly rely on text prompts, limiting their effectiveness for visually nuanced or ineffable styles that are difficult to articulate verbally. To bridge this gap, we propose Quiet Prompt (QuP), a reference-based concept erasure method that identifies complex concepts using fewer than ten reference images. QuP operates in two stages: (1) Concept Embedding Generation (CEG), which captures the intricate visual morphology of an unwanted concept into a single latent embedding; and (2) Concept leakage-aware Negative Guidance (CNG), which utilizes this embedding as a negative semantic condition to steer the diffusion process. In addition, to overcome the limitations of conventional fixed-scale negative guidance (negative prompting)—which often suffers from structural distortion and unintended content loss—we introduce concept leakage-aware negative guidance scale $n_g(t)$. This mechanism adaptively modulates the guidance strength at each denoising step based on a semantic measurement of concept leakage between the intermediate generated image and the target concept in a joint embedding space. Extensive evaluations on idiosyncratic art styles and object erasure tasks demonstrate that QuP achieves superior concept suppression and content preservation, outperforming state-of-the-art text-based baselines in both image fidelity and erasure precision. The source code will be publicly provided to ensure reproducibility of our work.

Modern video diffusion models generate increasingly realistic and temporally coherent videos, motivating their use as candidate world simulators. Yet it remains unclear whether these models internally encode physical structure, or merely reproduce motion patterns seen during training. We study this question by probing video diffusion models along latent trajectories corresponding to real videos with known physical plausibility. To obtain such trajectories, we approximately invert the deterministic sampling process by integrating the learned velocity field backward from a clean video latent to noise, giving access to the model’s intermediate states and attention maps. Using these recovered trajectories, we show that physical plausibility is linearly decodable from diffusion transformer states across IntPhys and InfLevel, reaching around 81\% average accuracy and outperforming dedicated representation-learning baselines such as V-JEPA and VideoMAE. Surprisingly, this signal is absent from the VAE latent input and emerges inside the denoising transformer itself, despite the model not being trained with a self-supervised predictive objective. These findings suggest that physically meaningful representations can arise as a byproduct of generative denoising.

Real-world datasets are inherently heterogeneous, yet how per-class structural differences and sampling imbalance shape the training dynamics of diffusion models—and potentially exacerbate disparities—remains poorly understood. While models typically transition from an initial phase of generalization to memorizing the training set, existing theory assumes homogeneous data, leaving open how class imbalance and heterogeneity reshape these dynamics. In this work, we develop a high-dimensional analytical framework to study class-dependent learning in score-based diffusion models. Analyzing a random-features model trained on Gaussian mixtures, we derive the feature-covariance spectrum to characterize per-class generalization and memorization times. We reveal the explicit hierarchy governing these dynamics: class variance is the primary determinant of learning order—consistently favoring higher-variance classes—while centroid geometry plays a secondary role. Sampling imbalance acts as a modulator that can reverse this ordering and, under strong imbalance, forces minority classes to acquire distinct, delayed speciation times during backward diffusion. Together, these results suggest that diffusion models can memorize some classes while others remain insufficiently learned. We validate our theoretical predictions empirically using U-Net models trained on Fashion MNIST.

Discrete diffusion models have emerged as strong alternatives to autoregressive language models, with recent multimodal work either finetuning unimodal diffusion bases or distilling autoregressive backbones for bi-modal generation. Diverging from these approaches, we introduce the first tri-modal \gls{mdm} \emph{pretrained from scratch} on text, image-text, and audio-text data, where all generation tasks are learned simultaneously within a single model. We systematically study the design space governing stability and efficiency at scale: we conduct the first empirical characterization of the critical batch size $B_{\text{crit}}$ under the SDE reparameterization of AdamW, and introduce a drift--horizon interpolation parameter $\gamma$ that balances gradient noise reduction and optimization horizon when scaling the token budget. We derive multimodal scaling laws and ablate modality mixing ratios, noise schedules, and anti-masking. Lastly, we pretrain a 3B model on 6.4T tokens, achieving competitive results in text generation, text-to-image, and text-to-speech.

Certifying authorities for safety-critical software audit recovered trace matrices module by module, not on average. Every existing distribution-free calibration delivers a marginal certificate, and we show the gap to the auditor's quantity is structural and unbounded. Under within-block exchangeability with cross-block heterogeneity in the conditional null score distribution, every procedure controlling marginal $\mathrm{FDR}$ at level $\alpha$ on a bipartite trace matrix with $K$ source modules admits worst-module false discovery proportion at least $1{-}\alpha{-}O(1/K)$ with probability at least $1{-}2/K$. We name this the block catastrophe and close it with $\mathrm{CERTRA}$, which wraps any black-box scorer, computes a within-block rank-indicator conformal $e$-value, and applies $e$-$\mathrm{BH}$ separately within each source module. The construction yields finite-sample per-module $\mathrm{FDR}$ control under arbitrary intra-module dependence, with matching power up to a rank-quantization factor. We release IndTrace-2026, a 12{,}406-link composite across aerospace, automotive, and medical-device specifications with professional ASIL/DAL/SIL annotation (Cohen's $\kappa{=}0.86$). $\mathrm{CERTRA}$ holds worst-block $\mathrm{FDP}$ at 0.118 where split-conformal drifts to 0.413, retains 0.741 of true links versus 0.572 for the strongest fair-operating-point baseline, and reduces audit deficiency 3.3-fold. Per-module distribution-free calibration is what an industrial audit can sign.

Training-free diffusion priors are powerful for inverse problems, but measurement guidance during reverse sampling is local: aggressive updates can improve immediate data fit while disrupting later denoising. We introduce a tethered predictive-inertial correction rule that separates proposal generation from acceptance. At each corrected step, the denoiser prediction anchors a frozen clean-space objective combining measurement fit with a noise-level-dependent tether. Heavy-ball dynamics with diffusion-scale predictive smoothing generate clean-state candidates; a separate verifier evaluates a finite dyadic set using the original unsmoothed objective, with the denoiser prediction retained as fallback. The accepted clean state is re-noised to continue sampling. The rule applies to pixel and latent diffusion with differentiable measurement operators, requires no retraining, and adds no denoiser calls inside the correction loop. We analyze when smoothing is inactive, how it changes nonlinear or decoder-composed proposal landscapes, and how objective verification yields local nonincrease, scale robustness, and displacement control. Across natural-image and accelerated MRI benchmarks, the method achieves competitive reconstruction quality with favorable speed--quality trade--offs, often reducing runtime relative to optimization-heavy diffusion solvers. On fastMRI knee reconstruction, it attains the strongest PSNR among compared methods at both acceleration factors.