This paper presents a novel generative framework for learning shared latent representations across multimodal data. Many advanced multimodal methods focus on capturing all combinations of modality-specific details across inputs, which can inadvertently obscure the high-level semantic concepts that are shared across modalities. Notably, Multimodal VAEs with low-dimensional latent variables are designed to capture shared representations, enabling various tasks such as joint multimodal synthesis and cross-modal inference. However, multimodal VAEs often struggle to design expressive joint variational posteriors and suffer from low-quality synthesis. In this work, ShaLa addresses these challenges by integrating a novel architectural inference model and a second-stage expressive diffusion prior, which not only facilitates effective inference of shared latent representation but also significantly improves the quality of downstream multimodal synthesis. We validate ShaLa extensively across multiple benchmarks, demonstrating superior coherence and synthesis quality compared to state-of-the-art multimodal VAEs. Furthermore, ShaLa scales to many more modalities while prior multimodal VAEs have fallen short in capturing the increasing complexity of the shared latent space.
论文检索
输入标题、作者或关键词，从 960 篇学术成果中精准定位
Graph Neural Networks (GNNs) have received increasing attention due to their ability to handle graph-structured data, yet their explainability remains a significant challenge. An effective solution is to provide the GNN models with counterfactual explanations, which aim to answer “How should the input instance be perturbed to change the model's prediction?". However, existing works mainly focus on generating explanations that can effectively alter model predictions, while neglecting whether the explanations remain aligned with the original data distribution, leading to the distribution shift problem. To address this problem, we propose a novel method called ICExplainer for generating explanations within the original distribution. Specifically, we introduce graph diffusion-based generative model into the counterfactual reasoning, treating it as an optimization objective for graph distribution learning. Taking insights from variational inference, we use it to estimate the true distribution of the input graphs to retain essential structural and semantic information. The inferred distribution is then utilized as prior knowledge to guide the reverse process, ensuring that generated explanations are both counterfactual and distributionally coherent. Extensive experiments conducted on both synthetic and real-world datasets demonstrate the superior performance of ICExplainer over existing methods.
Incorporating explicit reasoning rules within the latent space of language models (LMs) offers a promising pathway to enhance generalisation, interpretability, and controllability. While current Transformer-based language models have shown strong performance on Natural Language Inference (NLI) tasks, they often rely on memorisation rather than explicit rule-based generalisation. This work investigates how human-interpretable reasoning rules can be explicitly encoded within LMs with the support of Language Variational Autoencoders (VAEs), as a mechanism for generative control. We propose a complete pipeline for learning reasoning rules within Transformer-based language VAEs. This pipeline encompasses three rule-based reasoning tasks, a supporting theoretical framework, and a practical end-to-end architecture. The experiment illustrates the following findings: Disentangled reasoning: Under explicit signal supervision, reasoning rules (viewed as functional mappings) can be disentangled within the encoder’s parametric space. This separation results in distinct clustering of rules in the output feature space. Prior knowledge injection: injecting rule-based constraints into the Query enables the model to more effectively retrieve the stored Value from memory based on Key. This approach offers a simple method for integrating prior knowledge into decoder-only language models. Moreover, we found that FFN layers are better than attention layers at preserving the separation of reasoning rules in the model's parameters.
In robot manipulation, robot learning has become a prevailing approach. However, generative models within this field face a fundamental trade-off between the slow, iterative sampling of diffusion models and the architectural constraints of faster Flow-based methods, which often rely on explicit consistency losses. To address these limitations, we introduce MP1, which pairs 3D point-cloud inputs with the MeanFlow paradigm to generate action trajectories in one network function evaluation (1-NFE). By directly learning the interval-averaged velocity via the "MeanFlow Identity", our policy avoids any additional consistency constraints. This formulation eliminates numerical ODE-solver errors during inference, yielding more precise trajectories. MP1 further incorporates CFG for improved trajectory controllability while retaining 1-NFE inference without reintroducing structural constraints. Because subtle scene-context variations are critical for robot learning, especially in few-shot learning, we introduce a lightweight Dispersive Loss that repels state embeddings during training, boosting generalization without slowing inference. We validate our method on the Adroit and Meta-World benchmarks, as well as in real-world scenarios. Experimental results show MP1 achieves superior average task success rates, outperforming DP3 by 10.2% and FlowPolicy by 7.3%. Its average inference time is only 6.8 ms—19 times faster than DP3 and nearly 2 times faster than FlowPolicy.
Existing solutions for bundle recommendation (BR) have achieved remarkable effectiveness for predicting the user’s preference for prebuilt bundles. However, bundle-item (B-I) affiliation will vary dynamically in real scenarios. For ex ample, a bundle themed as ‘casual outfit’ may add ‘hat’ or remove ‘watch’ due to factors such as seasonal variations, changes in user preferences or inventory adjustments. Our empirical study demonstrates that the performance of main stream BR models may fluctuate or decline under item-level variability. This paper makes the first attempt to address the above problem and proposes Residual Diffusion for Bundle Recommendation (RDiffBR) as a model-agnostic generative framework which can assist a BR model in adapting this sce nario. During the initial training of the BR model, RDiffBR employs a residual diffusion model to process the item-level bundle embeddings which are generated by the BR model to represent bundle theme via a forward-reverse process. In the inference stage, RDiffBR reverses item-level bundle em beddings obtained by the well-trained bundle model under B-I variability scenarios to generate the effective item-level bundle embeddings. In particular, the residual connection in our residual approximator significantly enhances BR mod els’ ability to generate high-quality item-level bundle embed dings. Experiments on six BRmodelsandfourpublicdatasets from different domains show that RDiffBR improves the per formance of Recall and NDCG of backbone BR models by up to 23%, while only increases training time about 4%.
From Subtle to Significant: Prompt-Driven Self-Improving Optimization in Test-Time Graph OOD Detection
PDF ↗Graph Out-of-Distribution (OOD) detection aims to identify whether a test graph deviates from the distribution of graphs observed during training, which is critical for ensuring the reliability of Graph Neural Networks (GNNs) when deployed in open-world scenarios. Recent advances in graph OOD detection have focused on test-time training techniques that facilitate OOD detection without accessing potential supervisory information (e.g., training data). However, most of these methods employ a one-pass inference paradigm, which prevents them from progressively correcting erroneous predictions to amplify OOD signals. To this end, we propose a Self-Improving Graph Out-of-Distribution detector (SIGOOD), which is an unsupervised framework that integrates continuous self-learning with test-time training for effective graph OOD detection. Specifically, SIGOOD generates a prompt to construct a prompt-enhanced graph that amplifies potential OOD signals. To optimize prompts, SIGOOD introduces an Energy Preference Optimization (EPO) loss, which leverages energy variations between the original test graph and the prompt-enhanced graph. By iteratively optimizing the prompt by involving it into the detection model in a self-improving loop, the resulting optimal prompt-enhanced graph is ultimately used for OOD detection. Comprehensive evaluations on 21 real-world datasets confirm the effectiveness and outperformance of our SIGOOD method.
TOP-RL: Task-Optimized Progressive Token Pruning with Reinforcement Learning for Vision Language Models
PDF ↗In recent years, Large Vision-Language Models (LVLMs) have significantly advanced multimodal tasks. However, their inference requires intensive processing of numerous visual tokens and incurs substantial computational overhead. Existing methods typically compress visual tokens either at the input stage or in early model layers, ignoring variations across tasks and depths. To address these limitations, we introduce TOP-RL, a Task-Optimized Progressive token pruning framework based on Reinforcement Learning. TOP-RL formulates visual token pruning as a multi-stage Markov Decision Process (MDP). It employs an agent trained with dense and fine-grained reward signals to progressively generate differentiable binary masks. This enables TOP-RL to adaptively select crucial visual tokens tailored to each task, effectively balancing accuracy and computational efficiency. Extensive experiments on leading multimodal datasets and advanced LVLMs validate that TOP-RL effectively learns task-optimized pruning policies, significantly boosting inference efficiency while preserving robust performance. For instance, LLaVA-NeXT equipped with TOP-RL achieves a 1.9x speedup in inference time and a 9.3x reduction in FLOPs, with 96% performance preserved.
Among the diverse services provided by Location-Based Social Networks (LBSNs), Next Point-of-Interest (POI) recommendation plays a crucial role in inferring user preferences from historical check-in trajectories. However, existing sequential and graph-based methods frequently neglect significant mobility variations across distinct contextual scenarios (e.g., tourists versus locals). This oversight results in suboptimal performance due to two fundamental limitations: the inability to capture scenario-specific features and the failure to resolve inherent inter-scenario conflicts. To overcome these limitations, we propose the Multifaceted Scenario-Aware Hypergraph Learning method (MSAHG), a framework that adopts a scenario-splitting paradigm for next POI recommendation. Our main contributions are: (1) Construction of scenario-specific, multi-view disentangled sub-hypergraphs to capture distinct mobility patterns; (2) A parameter-splitting mechanism to adaptively resolve conflicting optimization directions across scenarios while preserving generalization capability. Extensive experiments on three real-world datasets demonstrate that MSAHG consistently outperforms five state-of-the-art methods across diverse scenarios, confirming its effectiveness in multi-scenario POI recommendation.
There is a growing demand for deploying large generative AI models on mobile devices. For recent popular video generative models, however, the Variational AutoEncoder (VAE) represents one of the major computational bottlenecks. Both large parameter sizes and mismatched kernels cause out-of-memory errors or extremely slow inference on mobile devices. To address this, we propose a low-cost solution that efficiently transfers widely used video VAEs to mobile devices. (1) We analyze redundancy in existing VAE architectures and get empirical design insights. By integrating 3D depthwise separable convolutions into our model, we significantly reduce the number of parameters. (2) We observe that the upsampling techniques in mainstream video VAEs are poorly suited to mobile hardware and form the main bottleneck. In response, we propose a decoupled 3D pixel shuffle scheme that slashes end-to-end delay. Building upon these, we develop a universal mobile-oriented VAE decoder, Turbo-VAED. (3) We propose an efficient VAE decoder training method. Since only the decoder is used during deployment, we distill it to Turbo-VAED instead of retraining the full VAE, enabling fast mobile adaptation with minimal performance loss. To our knowledge, our method enables real-time 720p video VAE decoding on mobile devices for the first time. This approach is widely applicable to most video VAEs. When integrated into four representative models, with training cost as low as $95, it accelerates original VAEs by up to 84.5× at 720p resolution on GPUs, uses as low as 17.5% of original parameter count, and retains 96.9% of the original reconstruction quality. Compared to mobile-optimized VAEs, Turbo-VAED achieves a 2.9× speedup in FPS and better reconstruction quality on the iPhone 16 Pro.
MCMoE: Completing Missing Modalities with Mixture of Experts for Incomplete Multimodal Action Quality Assessment
PDF ↗Multimodal Action Quality Assessment (AQA) has recently emerged as a promising paradigm. By leveraging complementary information across shared contextual cues, it enhances the discriminative evaluation of subtle intra-class variations in highly similar action sequences. However, partial modalities are frequently unavailable at the inference stage in reality. The absence of any modality often renders existing multimodal models inoperable. Furthermore, it triggers catastrophic performance degradation due to interruptions in cross-modal interactions. To address this issue, we propose a novel Missing Completion Framework with Mixture of Experts (MCMoE) that unifies unimodal and joint representation learning in single-stage training. Specifically, we propose an adaptive gated modality generator that dynamically fuses available information to reconstruct missing modalities. We then design modality experts to learn unimodal knowledge and dynamically mix the knowledge of all experts to extract cross-modal joint representations. With a mixture of experts, missing modalities are further refined and complemented. Finally, in the training phase, we mine the complete multimodal features and unimodal expert knowledge to guide modality generation and generation-based joint representation extraction. Extensive experiments demonstrate that our MCMoE achieves state-of-the-art results in both complete and incomplete multimodal learning on three public AQA benchmarks.
Real-time, high-fidelity monocular depth estimation from remote sensing imagery is crucial for numerous applications, yet existing methods face a stark trade-off between accuracy and efficiency. Although using Vision Transformer (ViT) backbones for dense prediction is fast, they often exhibit poor perceptual quality. Conversely, diffusion models offer high fidelity but at a prohibitive computational cost. To overcome these limitations, we propose Depth Detail Diffusion for Remote Sensing Monocular Depth Estimation (D³-RSMDE), an efficient framework designed to achieve an optimal balance between speed and quality. Our framework first leverages a ViT-based module to rapidly generate a high-quality preliminary depth map construction, which serves as a structural prior, effectively replacing the time-consuming initial structure generation stage of diffusion models. Based on this prior, we propose a Progressive Linear Blending Refinement (PLBR) strategy, which uses a lightweight U-Net to refine the details in only a few iterations. The entire refinement step operates efficiently in a compact latent space supported by a Variational Autoencoder (VAE). Extensive experiments demonstrate that D³-RSMDE achieves a notable 11.85% reduction in the Learned Perceptual Image Patch Similarity (LPIPS) perceptual metric over leading models like Marigold, while also achieving over a 40× speedup in inference and maintaining VRAM usage comparable to lightweight ViT models.
Diffusion models have been widely adopted in image and video generation. However, their complex network architecture leads to high inference overhead for its generation process. Existing diffusion quantization methods primarily focus on the quantization of the model structure while ignoring the impact of time-steps variation during sampling. At the same time, most current approaches fail to account for significant activations that cannot be eliminated, resulting in substantial performance degradation after quantization. To address these issues, we propose Time-Rotation Diffusion Quantization (TR-DQ), a novel quantization method incorporating time-step and rotation-based optimization. TR-DQ first divides the sampling process based on time-steps and applies a rotation matrix to smooth activations and weights dynamically. For different time-steps, a dedicated hyperparameter is introduced for adaptive timing modeling, which enables dynamic quantization across different time steps. Additionally, we also explore the compression potential of Classifier-Free Guidance (CFG-wise) to establish a foundation for subsequent work. TR-DQ achieves state-of-the-art (SOTA) performance on image generation and video generation tasks and a 1.38-1.89× speedup and 1.97-2.58× memory reduction in inference compared to existing quantization methods.
In recent years, neural image compression methods have achieved impressive performance in image compression tasks, most of which are based on variational auto-encoder with hyper-prior and autoregressive Gaussian entropy model. We first demonstrate that the way these end-to-end approaches handle quantization during training leads to a mismatch between the gradients direction of entropy model parameters (i.e., mean and standard deviation) and the direction they should be optimized towards during inference, making neural network difficult to learn accurate estimates of entropy model parameters. To address this issue, we then propose a two-step improvement: in the first step, use straight-through estimator to align the forward propagation during training with inference, thereby correcting the gradients of standard deviation parameters; in the second step, utilize gradients transfer that we propose and MSE-guided gradients to manually compensate for the gradients of mean parameters lost due to straight-through estimator. Finally, we also propose to freeze the auto-encoder and hyper auto-encoder in pre-trained models provided by existing works, and fine-tune only the modules that predict the entropy model parameters, enabling efficient validation of proposed improvements. Experimental results show that our improvements bring appreciable performance gains to state-of-the-art neural image compression models in recent years. Meanwhile, our improvements require no modification to the structure of pre-trained models and only lightweight fine-tuning, which shows strong plug-and-play capability and practical utility.
Augmentation-invariant Learning Strategy via Data Augmentation for Improving Model Generalization
PDF ↗Data augmentation is an effective technique for regularizing deep networks, which helps to enhance the generalizability and robustness of the model. However, in the field of medical imaging, traditional data augmentation techniques such as cropping, rotation, and degradation may inadvertently alter the critical characteristics of pathological lesions. Conventional semantic augmentation methods, such as altering the color and contrast of the object background, may also affect the structural features of medical images in uncontrolled semantic directions. Such operational conditions compromise the model's diagnostic reliability in medical contexts. To address this issue, we propose a surprisingly efficient implicit augmentation-invariant learning strategy (AILS) via variational Bayesian inference on differentially constrained feature manifolds. Parameterizing probability measures over tangent space through deep networks enables precise estimation of semantic direction distributions. Subsequently, geodesic-aware semantic features are sampled from the reparameterized variational posterior, achieving semantic-consistent augmentation. Simultaneously, to mine augmentation distribution invariance, we design the AiHLoss, which constrains the augmentation distribution to facilitate the network to learn augmentation invariance. Extensive experiments demonstrate that AILS exhibits high performance on public medical image datasets, outperforming existing augmentation methods.
Interactive segmentation aims to delineate a user-specified target in an image by leveraging positive and negative clicks. While effective on natural images, existing methods often fail in remote sensing scenarios, where satellite imagery is characterized by ultra-high resolution, sparse object distribution, and significant scale variation. These factors hinder accurate segmentation of fine-grained targets like roads, buildings, and aircraft. To overcome these problems, we propose CrossCut, a novel interactive segmentation framework tailored for remote sensing imagery. Unlike previous approaches that either process the entire image or treat each patch independently, CrossCut enables simultaneous segmentation across multiple patches by propagating user click information to all patches. This design allows the model to fully utilize click guidance regardless of object location, effectively resolving the challenge of inter-patch information isolation. Furthermore, CrossCut supports flexible inference by allowing segmentation results from different patch configurations to be fused, enhancing both accuracy and robustness. Extensive evaluations across multiple remote sensing datasets demonstrate that CrossCut achieves state-of-the-art performance. Quantitative results and visualizations show that CrossCut significantly advances the field of interactive segmentation for remote sensing imagery.
Numerous studies have demonstrated that Visual Question Answering (VQA) models are vulnerable to language priors and dataset biases, often leading to spurious correlations between questions and answers. As a result, these models excessively rely on linguistic cues, neglecting essential visual information and causing representational distortions. To address this challenge, we propose a novel Bayesian debiasing framework termed BayesVQA, which integrates three carefully designed mechanisms: Energy-guided Prior Variance (EPV), Energy-guided Posterior Sampling (EPS), and Energy-guided Likelihood Reweighting (ELR). Specifically, we explicitly decompose each sample's latent representation into a biased feature and a stochastic corrective perturbation δ. Using a Bayesian formulation, we model the posterior distribution of the perturbation δ conditioned on the predictive uncertainty, quantified via calibrated energy scores. To mitigate language bias, the posterior is optimized through energy-driven variational inference with an uncertainty-adaptive prior and sampling strategy. Moreover, the ELR mechanism incorporates an energy-based weighting of the reconstruction objective and enforces an energy-coherence constraint to emphasize challenging, high-uncertainty instances and align model confidence before and after debiasing. Extensive experiments conducted across multiple standard VQA benchmarks consistently validate the superior performance of our BayesVQA method over state-of-the-art competitors under distributional shifts and challenging bias conditions.
Reflection removal of a single image remains a highly challenging task due to the complex entanglement between target scenes and unwanted reflections. Despite significant progress, existing methods are hindered by the scarcity of high-quality, diverse data and insufficient restoration priors, resulting in limited generalization across various real-world scenarios. In this paper, we propose Dereflection Any Image, a comprehensive solution with an efficient data preparation pipeline and a generalizable model for robust reflection removal. First, we introduce a dataset named Diverse Reflection Removal (DRR) created by randomly rotating reflective mediums in target scenes, enabling variation of reflection angles and intensities, and setting a new benchmark in scale, quality, and diversity. Second, we propose a diffusion-based framework with one-step diffusion for deterministic outputs and fast inference. To ensure stable learning, we design a three-stage progressive training strategy, including reflection-invariant finetuning to encourage consistent outputs across varying reflection patterns that characterize our dataset. Extensive experiments show that our method achieves SOTA performance on both common benchmarks and challenging in-the-wild images, showing superior generalization across diverse real-world scenes.
Most 3D scene generation methods are limited to only generating object bounding box parameters while newer diffusion methods also generate class labels and latent features. Using object size or latent feature, they then retrieve objects from a predefined database. For complex scenes of varied, multi-categorical objects, diffusion-based latents cannot be effectively decoded by current autoencoders into the correct point cloud objects which agree with target classes. We introduce a Class-Partitioned Vector Quantized Variational Autoencoder (CPVQ-VAE) that is trained to effectively decode object latent features, by employing a pioneering class-partitioned codebook where codevectors are labeled by class. To address the problem of codebook collapse, we propose a class-aware running average update which reinitializes dead codevectors within each partition. During inference, object features and class labels, both generated by a Latent-space Flow Matching Model (LFMM) designed specifically for scene generation, are consumed by the CPVQ-VAE. The CPVQ-VAE's class-aware inverse look-up then maps generated latents to codebook entries that are decoded to class-specific point cloud shapes. Thereby, we achieve pure point cloud generation without relying on an external objects database for retrieval. Extensive experiments reveal that our method reliably recovers plausible point cloud scenes, with up to 70.4% and 72.3% reduction in Chamfer and Point2Mesh errors on complex living room scenes.
The generalization capability of deepfake detectors is crucial for real-world applications. Data augmentation to generate synthetic fake faces has served as an effective strategy to enhance generalization. Interestingly, current state-of-the-art (SoTA) methods rely on fixed augmentation strategies, raising a fundamental question: Can a single static augmentation approach suffice, or does the diversity of forgery features necessitate dynamic strategies? We argue that existing methods overlook the evolving complexity of real-world forgery patterns, such as facial warping, expression manipulation, and compression artifacts, which cannot be fully simulated by fixed policies. To bridge this gap, we propose CRDA (Curriculum Reinforcement-Learning Data Augmentation), a novel framework that guides the detector to progressively master multi-domain forgery features from simple to complex. CRDA synthesizes augmented samples using a configurable pool of forgery operations and dynamically generates adversarial samples tailored to the detector’s current learning state. Key to our approach is the integration of reinforcement learning (RL) and causal inference. To efficiently explore the vast augmentation space, an RL agent dynamically selects augmentation actions based on the detector’s performance, ensuring continuous adaptation to increasingly challenging forgeries. Simultaneously, the agent’s output is designed to introduce variations in action spaces, generating heterogeneous forgery patterns. These variations are guided by causal inference theory, which mitigates spurious correlations by suppressing task-irrelevant biases and enforcing the model to focus on causally invariant features. This integration ensures robust generalization by decoupling synthetic augmentation patterns from the model’s learned representations. Extensive experiments demonstrate that the proposed method significantly improves the generalizability of the detector, achieving superior performance compared to state-of-the-art methods on multiple cross-domain datasets.
VPSentry: Semi-supervised Video Polyp Segmentation via Sentry-guided Long-term Prototype Fusion with Correlation Dynamic Propagation
PDF ↗Automated polyp segmentation in colonoscopy videos is an essential computer-aided technology for early detection and removal of polyps. However, most existing video polyp segmentation methods are designed with pixel-level temporal learning mechanisms, at the cost of time-consuming frame-wise annotations. In this paper, we present VPSentry, a novel semi-supervised segmentation model with a sentry mechanism. Our model integrates a prototype memory to store the long-term spatiotemporal cues of colonoscopy videos. Moreover, we devise adaptive prototypes to capture and generalize critical representations from individual frames, enabling long-term temporal fusion across labeled and unlabeled frames. In addition, we propose a correlation dynamic propagation module that propagates information from prototypes to features while simultaneously extracting dynamic features to perceive variations in polyp details between adjacent frames. Since colonoscopy scenes may change among consecutive frames, we further employ a sentry mechanism to assess the inter-frame continuity. This mechanism guides the prototype memory updating and the correlation dynamic propagation, further facilitating robust temporal propagation and dynamic detail perception for semi-supervised learning of long-term colonoscopy video sequences. Extensive experiments on the large-scale SUN-SEG dataset demonstrate that our model achieves optimal segmentation performance with real-time inference efficiency.