NeurIPS 2026 Papers

What does it mean to create a new concept, rather than retrieve a familiar one? Diffusion models sampled many times at the same prompt vary only within the narrow stylistic envelope set by a fixed prompt embedding. We propose an operational definition of creativity as the generation of concepts that are initially unfamiliar yet rapidly learnable to an adaptive observer, and formalize it as a bilevel optimization between a Creator that generates and an Appraiser that adapts: the Appraiser's improvement under a brief inner-loop adaptation provides the reward signal that the Creator maximizes. We call this framework CAMEL (Creator-Appraiser Meta-Learning), and instantiate it on MNIST with an autoencoder Appraiser and on natural images with a CLIP Appraiser based on a low-rank adapter over the text projection. CAMEL produces outputs that lie outside the basin reachable by vanilla diffusion sampling at any sample count, while remaining recognizable as the prompt's class to an off-the-shelf observer.

Diffusion-based video generation has achieved impressive visual fidelity, yet enforcing temporally consistent motion under explicit control signals remains challenging. Recent noise-warping approaches construct temporally correlated noise by transporting the initial latent noise via diffeomorphic deformation fields derived from optical flow, but this spatial resampling inherently destroys the i.i.d. Gaussian structure of the noise, requiring post-hoc correction steps at the cost of non-trivial computational overhead. We propose \textbf{Seed-Your-Motion}, by conditioning the a priori latent noise on optical flow via per-location orthogonal transformations in latent channel space, parameterized as products of the Householder Transformation. This channel-space formulation has an advantage of preserving the spatial Gaussianity exactly by construction, requiring neither Jacobian corrections nor post-hoc distribution restoration, while at the same time inducing structured cross-frame correlations that ensures geometric consistency across time. With no learnable parameters and a single forward pass, our method is agnostic to model architecture and can be plugged seamlessly into any video diffusion fine-tuning pipeline. Extensive experiments on motion-transfer and camera-controlled video generation benchmarks demonstrate that our method matches or outperform the state-of-the-art noise-warping baseline, while achieving a 2.7× speedup in processing time. The results suggest that the channel-space Householder transformations provide a principled and efficient alternative to noise-warping based motion-controllable video diffusions.

The rapid emergence of generative image models has led to the development of specialized watermarking techniques, particularly in-generation methods such as seed-based embedding. However, current evaluations in this area remain largely empirical, making them heavily reliant on the specific model architectures used for generation and inversion. This prevents any clear conclusion on the performance of any method, especially regarding security, for which a rigorous definition is lacking. Against this approach, we argue that the effectiveness of a watermarking scheme should be established purely through a thorough theoretical analysis. This is enabled by decoupling the model-dependent part from the actual decision mechanism of the watermarking system. Using this decoupling, we introduce a formal evaluation framework based on security, robustness, and fidelity. This allows precise comparisons between watermarking systems through a characteristic surface representing the trade-off between these three quantities, independent of any generative model. Based on this framework, we propose SSB, a novel watermarking method that generalizes previous seed-based methods by allowing to reach any security-robustness-fidelity regime on its characteristic surface. This work opens the door to the design of modern watermarking systems with theoretical guarantees that do not necessitate any costly empirical evaluations.

Discrete diffusion models generate sequences by iteratively denoising samples corrupted by categorical noise, offering an appealing alternative to autoregressive decoding for structured and symbolic generation. However, standard training targets a likelihood-based objective that primarily matches the data distribution and provides no native mechanism for enforcing hard constraints or optimizing non-differentiable properties at inference time. This work addresses this limitation and introduces Search-Augmented Masked Diffusion (SearchDiff), a training-free neurosymbolic inference framework that integrates informed search directly into the reverse denoising process. At each denoising step, the model predictions define a proposal set that is optimized under a user-specified property satisfaction, yielding a modified reverse transition that steers sampling toward probable and feasible solutions. Experiments in biological design and symbolic reasoning illustrate that SearchDiff substantially improves constraint satisfaction and property adherence, while consistently outperforming discrete diffusion and autoregressive baselines.

Diffusion models have become the dominant paradigm in text-to-image generation, and test-time scaling (TTS) improves sample quality by allocating additional computation at inference. Existing TTS methods, however, resample the entire image, while generation quality is often spatially heterogeneous. This leads to unnecessary computation on regions that are already correct, and localized defects remain insufficiently corrected. In this paper, we explore a new direction -- Localized TTS -- that adaptively resamples defective regions while preserving high-quality regions, thereby substantially reducing the search space. This raises two challenges: accurately localizing defects and maintaining global consistency. We propose $\textbf{LoTTS}$, the first fully training-free framework for localized TTS. For defect localization, LoTTS contrasts cross-/self-attention signals under quality-aware prompts (e.g., ''high-quality'' vs. ''low-quality'') to identify defective regions, and then refines them into coherent masks. For consistency, LoTTS perturbs only defective regions and denoises them locally, ensuring that corrections remain confined while the rest of the image remains undisturbed. Extensive experiments on SD2.1, SDXL, and FLUX demonstrate that LoTTS achieves state-of-the-art performance:~it consistently improves both local quality and global fidelity, while reducing GPU cost by $2$-$4\times$ compared to Best-of-$N$ sampling. These findings establish localized TTS as a promising new direction for scaling diffusion models at inference time.

Guard models are the last line of defense between a language model and a harmful output, yet their training objective is surprisingly narrow. Existing guards learn to predict a single verdict token from a conversational context, concentrating supervision on a single target. The consequences are structural: models latch onto shortcut features, are overconfident, and remain sensitive to where safety evidence appears in the sequence rather than its role in the full context. We propose a different framing. Rather than predicting a label from text, our $\textbf{LLaDA Guard}$ asks which label better explains the text: scoring the prompt or response under each label hypothesis and classifying based on their difference. This shifts supervision to every token in the moderated region, forcing the model to account for full content rather than its most discriminative fragments. We instantiate this idea with a masked diffusion language model, fine-tuning LLaDA-8B-Instruct with a class-conditional reconstruction objective using LoRA and requiring no architectural changes beyond the base model. LLaDA Guard leads on average rank against discriminative baselines trained on stronger backbones across seven held-out safety benchmarks, while exhibiting substantially better confidence calibration (ECE 0.0875 vs. 0.1384 for Qwen3Guard), less over-defense on benign prompts with $\texttt{unsafe}$-looking cues, and less prompt leakage when moderating responses. Its generative nature further enables token-level risk localization as a natural byproduct, yielding a pipeline for rewriting $\texttt{unsafe}$ prompts into $\texttt{safe}$ equivalents without additional training and achieving a 60.7% average conversion-to-$\texttt{safe}$ rate. The weights and code for reproduction are available at the following [anonymized repository](https://anonymous.4open.science/r/LLaDA-Guard2026-B056/).

Online safe reinforcement learning (RL) seeks policies that maximize reward while satisfying safety constraints. A popular line of research in safe RL relaxes safety to a soft expected-cost constraint and solves the resulting Constrained Markov Decision Process (CMDP) via primal-dual Lagrangian updates that only enforce safety on average. To address this limitation, hard, state-wise constraints are introduced and often imposed through Hamilton-Jacobi (HJ) reachability. Yet such constraints require solving different objectives in the feasible and infeasible regions of the state space: reward maximization in the former, recovery toward the feasible regions in the latter. The resulting target action distributions are inherently multimodal, and this structure poses a fundamental challenge for the Gaussian or deterministic actors used in existing HJ-based safe RL, which often collapse onto suboptimal modes. Diffusion policies provide the expressiveness needed to represent such distributions, and recent work on Q-score matching offers a route to training them for online RL by score regression---but has been applied only to reward maximization. Building on this framework, we propose Safe Score Matching (SSM), an off-policy actor-critic method that adapts Q-score matching to hard-constrained safe RL by gating a two-branch score target with HJ reachability: inside the feasible set, the denoising process degenerates to Q-score matching on actions classified as viable by the HJ critic, encouraging reward maximization; outside, a recovery branch biases denoising toward regions with lower worst-case violation, guided by the negative action gradient of the HJ safety value. Across high-dimensional quadrotor and fixed-wing trajectory-tracking and stabilize-and-avoid benchmarks, SSM achieves competitive performance without sacrificing safety, in contrast to primal-dual and reachability-based baselines, which tend to trade off performance for safety and are thus overly conservative.

Flow matching has recently emerged as a strong paradigm for state-of-the-art text-to-image (T2I) generation, enabling high-quality generation with a small number of sampling steps. As these models are increasingly integrated into real-world applications, ensuring safe and non-sensitive content generation has become a critical requirement. However, adapting safety and concept removal methods to this new generation framework remains an open challenge. Specifically, prior methods largely rely on iterative trajectory steering across a number of denoising steps or on CLIP-centeric prompt embedding manipulation. These design assumptions pose fundamental bottlenecks for safety in flow matching-based T2I generation, where limited sampling steps constrain iterative correction and modern context-aware text encoders diminish the effectiveness of embedding-level interventions. In this paper, we propose VESFlow, a training-free safety method tailored to flow matching with extremely few sampling steps. Leveraging the fact that flow matching models learn the marginal velocity (or average velocity in MeanFlow), we directly edit the velocity field via a Bayesian decomposition of the safe-conditional posterior. VESFlow steers the trajectory toward safe outputs while leaving the conditioning prompt unchanged. Building on the observation that VESFlow leaves outputs unchanged under benign prompts, we further introduce a risk score filtering that bypasses velocity editing to reduce computational cost while preserving benign prompt generation. Based on this filtering, we proposed VESFlow+ which provides stronger safety protection when filtered by the risk score. Experimental results show that our method removes the target concept, reducing the detection rate by NudeNet to 6.3\% for step-4 model, while preserving fidelity on benign prompts. Code is available at the supplementary file.

We study the problem of human gaze modeling, which aims to generate the gaze patterns a viewer produces while observing a visual stimulus. Gaze is primarily captured through two modalities: continuous eye-tracking trajectories, which describe fine-grained motion dynamics, and discrete scanpaths, which describe high-level fixation structure. Because gaze varies substantially across viewers and trials, we treat this variability as a defining property rather than noise and model gaze as a stochastic generative process. Existing generative gaze models supervise on only one of these two representations in isolation. We hypothesize that trajectories and scanpaths describe gaze at complementary scales and are jointly informative during training, and test this hypothesis through ST-DiffEye, a joint trajectory-scanpath diffusion framework that couples both modalities by concatenating them as an additional raw input channel, requiring no architectural overhead beyond an input and output channel expansion. We further introduce a principled evaluation framework based on the Continuous Ranked Probability Score (CRPS), which generalizes any existing sequence similarity metric into a proper scoring rule that jointly assesses the accuracy and diversity of generated gaze. Experiments on task-driven visual search, covering both target-present and target-absent scenarios, and on free-viewing benchmarks demonstrate state-of-the-art performance. These results, along with detailed ablations, confirm the benefit of joint modeling and the value of distribution-aware evaluation in capturing the intrinsic variability of human gaze.

Autoregressive language models generate under a fixed left-to-right order, whereas diffusion language models (dLLMs) denoise masked tokens through flexible any-order trajectories that may commit multiple positions per step. This makes dLLMs a distinctive setting for test-time scaling: additional inference compute can diversify both token choices and the order in which positions are committed. Existing samplers do not explicitly allocate these two sources of diversity. Confidence-based decoding is reliable but often redundant, while global temperature sampling and random remasking inject stochasticity without distinguishing useful exploration from unreliable commitments. We propose SIGMA, a training-free sampler that formulates each denoising step as a constrained allocation problem over decoding efficiency, token-level diversity, trajectory-level diversity, and model-internal commitment uncertainty. The resulting objective decomposes joint sampling entropy into token and trajectory terms, yielding a sigmoid-form stochastic gate for position selection and an adaptive per-position temperature for token sampling. Experiments on LLaDA and Dream across math and code benchmarks show that SIGMA improves the accuracy--cost Pareto frontier and can be integrated into existing dLLM test-time scaling pipelines.

Diffusion transformers (DiTs) have emerged as a dominant architecture for text-to-image generation, yet their performance drops when generating at resolutions beyond their training range. Existing training-free approaches mitigate this by modifying inference-time attention behavior, often through Rotary Position Embeddings (RoPE) extrapolation combined with attention scaling. However, these strategies apply a uniform and content-agnostic scaling across RoPE components with distinct frequency characteristics, inducing a trade-off between preserving global structure and recovering fine detail. We introduce \textbf{SEGA}, a training-free method that dynamically scales attention across RoPE components according to the latent’s spatial-frequency structure at each denoising step. This adaptive scaling improves both structural coherence and fine-detail fidelity. Experiments show that SEGA consistently improves high-resolution synthesis across multiple target resolutions, outperforming state-of-the-art training-free baselines.

Vector quantization (VQ) with autoregressive (AR) token modeling is a widely adopted and highly competitive paradigm for time-series generation. However, such models are fundamentally limited by exposure bias: during inference, errors can accumulate across sequential predictions, leading to pronounced quality degradation in long-horizon generation. To address this, we propose **SDFlow** (**S**imilarity-**D**riven **Flow** Matching), a non-autoregressive framework that operates entirely in the frozen VQ latent space and enables parallel sequence generation via flow matching. We tackle three key challenges in making this transition: (1) eliminating exposure bias by replacing step-wise token prediction with a global transport map; (2) mitigating the high-dimensionality of VQ token spaces via a low-rank manifold decomposition with a learned anchor prior over the latent manifold; and (3) incorporating discrete supervision into continuous transport dynamics by introducing a categorical posterior over codebook indices within a variational flow-matching formulation. Extensive experiments show that SDFlow achieves state-of-the-art performance, improving Discriminative Score and substantially reducing Context-FID, particularly for challenging long-sequence generation. Moreover, SDFlow provides significant inference speedups over autoregressive baselines, offering both high fidelity and computational efficiency. Code is available at *https://anonymous.4open.science/r/SDFlow-D6F3/*

We propose SCULPT, a state-of-the Art masked discrete diffusion model (MDMs) for high resolution text-to-image synthesis. Compared with prior works on masked image generation, SCULPT addresses two key challenges. First, unlike continuously diffusion models which progressively refines image latent across the entire image, vanilla MDMs do not have self-correcting capability because discrete tokens cannot be changed once it's unmasked. Second, while scaling the vocabulary size of discrete image tokenizers can improve reconstruction quality, it also introduces optimization challenges for training generative models as per-token training signal becomes more spares. To address the first challenge, SCULPT incroprates a token-editing mechanism where the model can dynamically correct already-unmasked output tokens during inference. To address the second challenge. we proposed a Goruped Corss Entrophy (GCE) objective that assigns positive learning signal to adjacent tokens of the ground truth in the embedding space. To improve the training efficiency, we further implemented a custom fused operator that greatly reduce the VRAM requirements during training in large-vocabulary setting. Experiment results show that these innovations significant improves the training efficiency and image fidelity of masked discrete image generators.

Diffusion models achieve strong performance across image and audio synthesis, but their inference remains computationally expensive, requiring tens to hundreds of sequential evaluations of a large denoising network. We observe that the effective rank of the denoiser output varies sharply across timesteps and architectures, motivating step- and block-wise adaptation of computation. We propose Step-Adaptive Rank Adjustment (SARA), a training-free framework that treats the SVD rank of pretrained linear layers as an inference-time control variable, adapted per step and block by SARA-DP and per step by SARA-Online. SARA introduces the Low-Rank Approximation Error (LRAE) — the discrepancy a rank-r truncation induces against the full-rank network — and instantiates it in two algorithms: SARA-DP measures LRAE on a few warmup prompts and solves the globally optimal rank assignment under a MACs budget by dynamic programming; SARA-Online computes a closed-form, spectrum-based LRAE from the model output at inference time, requiring no calibration and no trajectory lookahead. SARA applies uniformly to UNet (SDXL), MMDiT (SD3), DiT (Stable Audio Open), and hybrid MMDiT/DiT (TangoFlux) backbones, spanning ϵ, v, and rectified-flow parameterizations. SARA combines with existing solver- and cache-axis acceleration to yield compounded wall-clock speedups in our experiments. Together, these results introduce rank as a new axis of diffusion inference acceleration.

While dataset distillation has witnessed significant advances in image classification, its extension to dense prediction tasks such as semantic segmentation is plagued by two core bottlenecks: bi-level optimization-based approaches suffer from prohibitive computational costs stemming from pixel-wise gradient unrolling, whereas proxy-based generative methods merely infuse visual textures into fixed semantic masks derived from real data, thus failing to condense rich semantic knowledge into a compact set of informative, novel scene generations. To address both bottlenecks simultaneously, we propose a novel Sample-Set Reinforcement Learning (S$^2$-RL) framework for generative semantic segmentation dataset distillation (SSDD). Specifically, S$^2$-RL formulates SSDD as a diffusion model based text-to-image generation paradigm, enabling flexible generation of images with informative, novel semantic distributions via a concise text prompt. Furthermore, we fine-tune a diffusion policy via GRPO, leveraging a sample-set dual reward paradigm: for individual samples, it enforces semantic alignment with input text prompts and intra-class feature diversity; for sample sets, it maximizes inter-sample semantic distribution diversity by solving a contextual multi-armed bandit problem. These design choices enable S$^2$-RL to distill large-scale semantic segmentation datasets into a compact set of generated samples that encapsulate diverse, informative semantic knowledge without the need of pixel-wise gradient unrolling. Extensive experiments on the ADE20K and COCO datasets demonstrate that S$^2$-RL achieves substantial and consistent improvements over state-of-the-art baselines, establishing a new benchmark for SSDD.

Medication recommendation requires generating drug combinations that are therapeutically effective while minimizing harmful drug-drug interactions (DDIs). Both objectives are rooted in the combinatorial nature of prescriptions. Existing discriminative methods predict drugs independently, neglecting inter-drug dependencies; autoregressive methods introduce sequential dependencies but impose arbitrary generation orders and accumulate errors. Both paradigms confine DDI mitigation to training-time penalties, offering limited capacity to regulate interactions during inference for individual prescriptions. We reformulate medication recommendation as a discrete diffusion process and propose RxDiff, which leverages iterative denoising for bidirectional and revisable generation and provides a natural interface for dynamic DDI regulation at each step. A clinically grounded dual-guided forward process and an inference-time interaction-aware modulation mechanism are designed to jointly enable RxDiff to achieve consistent accuracy improvements, lower and more individually controlled DDI rates, and robustness to evolving DDI knowledge across MIMIC-III, MIMIC-IV, and a real-world inpatient dataset. Code is available at https://anonymous.4open.science/r/RxDiff-9157.

Diffusion models have emerged as a powerful tool across diverse domains. However, their purely data-driven nature can produce samples that violate domain-governing constraints. We introduce a plug-and-play Reinforcement Learning framework that optimizes initial noise samples in the latent space of frozen, pre-trained diffusion models. Leveraging the near-spherical geometry of high-dimensional Gaussian distributions, we introduce a novel rotation-matrix-based scheme for efficient latent space exploration. This steers the model toward more feature-preserving outputs, guided by task-specific rewards. We evaluate our method on three diffusion models: one trained on solutions of the Darcy Flow PDE, one on a synthetic dataset with complex structural features, and a text-conditioned one. Across all three settings, our framework yields significant improvements in sample quality, achieving a ${\sim}25\\%$ relative reduction in PDE residual, up to a ${\sim}44\\%$ relative improvement on the synthetic dataset's feature-alignment metric, and up to a ${\sim}80\\%$ relative improvement on human preference, compared to the vanilla diffusion models. Finally, we show that rotation-matrix-based exploration significantly outperforms unconstrained exploration, validating our geometry-aware approach and establishing a more effective method for latent space control.

Latent Action Models (LAMs) have emerged as an effective paradigm for handling heterogeneous datasets during Vision-Language-Action (VLA) model pretraining, offering a unified action space across embodiments. However, existing LAMs often rely on discrete quantization encode and decode pipelines, which can lead to trivial frame reconstruction behavior, limited representational capacity, and a lack of physically meaningful structure. We introduce RotVLA, a VLA framework built on a continuous rotational latent action representation. Latent actions are modeled as elements of ${\rm SO}(n)$, providing continuity, compositionality, and structured geometry aligned with real-world action dynamics. A triplet frame learning framework further enforces meaningful temporal dynamics while avoiding degeneration. RotVLA consists of a VLM backbone and a flow-matching action head, pretrained on large-scale cross-embodiment robotic datasets and human videos with latent-action supervision. For downstream robot control, the flow-matching head is extended into a unified action expert that jointly denoises latent and robot actions. Here, latent actions serve as a latent planner, providing high-level guidance that conditions action generation. With only 1.7B parameters and 1700+ hours of pretraining data, RotVLA achieves 98.2\% on LIBERO and 89.6\% / 88.5\% on RoboTwin2.0 under clean and randomized settings, respectively. It also demonstrates strong real-world performance on manipulation tasks, consistently outperforming existing VLA models.

A core challenge in structural biophysics is generating biomolecular conformations that are both physically plausible and consistent with experimental measurements. While sequence-to-structure diffusion models provide powerful priors, posterior sampling methods steer generation by perturbing atomic coordinates with gradients from experimental likelihoods. However, when the target lies in a low-density region of the prior, these methods require aggressive upweighting of the likelihood that can destabilize sampling and be sensitive to hyperparameters. We propose EmbedOpt, an inference-time steering framework that introduces an orthogonal optimization axis: rather than performing posterior sampling under a fixed prior, EmbedOpt directly optimizes the prior by updating the model's conditional embedding. This embedding space encodes rich coevolutionary signals, so optimizing it shifts the structural prior to align with experimental constraints. Empirically, EmbedOpt matches coordinate-based posterior sampling baselines on sparse distance constraints and outperforms them on cryo-electron microscopy map fitting, including real, noisy experimental ones. Furthermore, EmbedOpt's smooth optimization behavior yields robustness to hyperparameters spanning two orders of magnitude and enables comparable performance with fewer diffusion steps.