Weight Averaging (WA) has emerged as a powerful technique for enhancing generalization by promoting convergence to a flat loss landscape, which correlates with stronger out-of-distribution performance. However, applying WA directly to multi-modal domain generalization (MMDG) is challenging: differences in optimization speed across modalities lead WA to overfit to faster-converging ones in early stages, suppressing the contribution of slower yet complementary modalities, thereby hindering effective modality fusion and skewing the loss surface toward sharper, less generalizable minima. To address this issue, we propose MBCD, a unified collaborative distillation framework that retains WA's flatness-inducing advantages while overcoming its shortcomings in multi-modal contexts. MBCD begins with adaptive modality dropout in the student model to curb early-stage bias toward dominant modalities. A gradient consistency constraint then aligns learning signals between uni-modal branches and the fused representation, encouraging coordinated and smoother optimization. Finally, a WA-based teacher conducts cross-modal distillation by transferring fused knowledge to each uni-modal branch, which strengthens cross-modal interactions and steer convergence toward flatter solutions. Extensive experiments on MMDG benchmarks show that MBCD consistently outperforms existing methods, achieving superior accuracy and robustness across diverse unseen domains.
论文检索
输入标题、作者或关键词，从 11,272 篇学术成果中精准定位
Driving world models are used to simulate futures by video generation based on the condition of the current state and actions. However, current models often suffer serious error accumulations when predicting the long-term future, which limits practical applications. Recent studies utilize the Diffusion Transformer (DiT) as the backbone of driving world models to improve learning flexibility. However, these models are always trained on short video clips, and multiple roll-out generations struggle to produce consistent and reasonable long videos due to the training-inference gap. To this end, we propose several solutions to build a simple yet effective long-term driving world model. First, we hierarchically decouple world model learning into large motion learning and bidirectional continuous motion learning. Then, considering the continuity of driving scenes, we propose a simple distillation method where fine-grained video flows are self-supervised signals for coarse-grained flows. The distillation is designed to improve the coherence of infinite video generation. The coarse-grained and fine-grained modules are coordinated to generate long-term and temporally coherent videos. On NuScenes, compared with the state-of-the-art front-view models, our model improves FVD by 27% and reduces inference time by 85% for the video task of generating 110+ frames.
The development of multimodal large language models (MLLMs) has advanced general video understanding. However, existing video evaluation benchmarks primarily focus on non-interactive videos, such as movies and recordings. To fill this gap, this paper proposes the first omnimodal benchmark for interactive livestream videos, LiViBench. It features a diverse set of 24 tasks, highlighting the perceptual, reasoning, and livestream-specific challenges. To efficiently construct the dataset, we design a standardized semi-automatic annotation workflow that incorporates the human-in-the-loop at multiple stages. The workflow leverages multiple MLLMs to form a multi-agent system for comprehensive video description and uses a seed-question-driven method to construct high-quality annotations. All interactive videos in the benchmark include audio, speech, and real-time comments modalities. To enhance models' understanding of interactive videos, we design tailored two-stage instruction-tuning and propose a Video-to-Comment Retrieval (VCR) module to improve the model's ability to utilize real-time comments. Based on these advancements, we develop LiVi-LLM-7B, an MLLM with enhanced knowledge of interactive livestreams. Experiments show that our model outperforms larger open-source models with up to 72B parameters, narrows the gap with leading proprietary models on LiViBench, and achieves enhanced performance on general video benchmarks, including VideoMME, LongVideoBench, MLVU, and VideoEval-Pro.
Zero-shot classifier expansion aims to adapt existing model to new, unseen classes. It utilizes class attributes or textual descriptions to learn a mapping from the semantic space to the classifier's weight space, without requiring new visual training data. However, the learning process for this mapping relies solely on correlating semantic patterns with their corresponding classifier weights and lacks explicit modeling of inter-class differences. This makes it difficult for the model to capture the critical discriminative features required to define classification boundaries. To overcome this limitation, we reframe the problem from a causal perspective and introduce a novel framework driven by counterfactuals. Our method first generates factual descriptions alongside corresponding inter-class counterfactuals to pinpoint the causal attributes essential for classification, then refines these representations via a mutual purification process, and finally leverages a novel separation loss to explicitly push the factual and counterfactual classifier weights apart. This strategy forces the model to forge clearer and more discriminative classification boundaries, achieving more accurate and robust classification. Extensive experiments demonstrate that our approach significantly outperforms existing state-of-the-art methods.
We study experimentation under endogenous network interference. Interference patterns are mediated by an endogenous graph, where edges can be formed or eliminated as a result of treatment. We show that conventional estimators are biased in these circumstances, and present a class of unbiased, consistent and asymptotically normal estimators of total treatment effects in the presence of such interference. We show via simulation that our estimator outperforms existing estimators in the literature. Our results apply both to bipartite experimentation, in which the units of analysis and measurement differ, and the standard network experimentation case, in which they are the same.
The Area Under the ROC Curve (AUC) is an important evaluation metric for both linear and, in particular, nonlinear classification models, owing to its robustness against class imbalance. Sparse learning with an ℓ₀ constraint can enhance model interpretability and generalization. Prior work has shown that, in the linear setting, the pairwise formulation of AUC maximization can be reformulated as a standard pointwise empirical risk minimization problem, which enables efficient optimization using hard-thresholding gradient descent for ℓ₀-constrained AUC maximization. Extending this approach to the nonlinear setting remains largely unexplored, even though we establish that pairwise AUC maximization in this setting is equivalent to a pointwise compositional optimization problem; however, designing a compositional optimization algorithm compatible with hard-thresholding operators remains an open challenge. To address this challenge, in this paper, we propose a novel algorithm—Compositional Stochastic Hard Thresholding (CSHT)—for nonlinear sparse AUC maximization. Specifically, CSHT integrates stochastic variance-reduced gradient techniques with hard-thresholding projections to effectively reduce gradient estimation variance while enforcing sparsity. Notably, we provide a rigorous convergence analysis and prove that CSHT achieves linear convergence up to a tolerance bound. To the best of our knowledge, this is the first stochastic hard-thresholding algorithm tailored for nonlinear sparse AUC maximization. Extensive experiments on (a) nonlinear sparse AUC maximization using Random Fourier Feature-based kernel approximation and (b) universal adversarial attack scenarios demonstrate the superior performance of CSHT over existing methods, attributed to its unified treatment of nonlinearity and sparsity.
Long-term Time Series Forecasting is crucial across numerous critical domains, yet its accuracy remains fundamentally constrained by the receptive field bottleneck in existing models. Mainstream Transformer- and Multi-layer Perceptron (MLP)-based methods mainly rely on finite look-back windows, limiting their ability to model long-term dependencies and hurting forecasting performance. Naively extending the look-back window proves ineffective, as it not only introduces prohibitive computational complexity, but also drowns vital long-term dependencies in historical noise. To address these challenges, we propose CometNet, a novel Contextual Motif-guided Long-term Time Series Forecasting framework. CometNet first introduces a Contextual Motif Extraction module that identifies recurrent, dominant contextual motifs from complex historical sequences, providing extensive temporal dependencies far exceeding limited look-back windows; Subsequently, a Motif-guided Forecasting module is proposed, which integrates the extracted dominant motifs into forecasting. By dynamically mapping the look-back window to its relevant motifs, CometNet effectively harnesses their contextual information to strengthen long-term forecasting capability. Extensive experimental results on eight real-world datasets have demonstrated that CometNet significantly outperforms current state-of-the-art (SOTA) methods, particularly on extended forecast horizons.
Training of large-scale models is both computationally intensive and often constrained by the availability of labeled data. Model merging offers a compelling alternative by directly integrating the weights of multiple source models without requiring additional data or extensive training. However, conventional model merging techniques, such as parameter averaging, often suffer from the unintended combination of non-generalizable features, especially when source models exhibit significant weight disparities. Comparatively, model ensembling generally provides more stable and superior performance that aggregates multiple models by averaging outputs. However, it incurs higher inference costs and increased storage requirements. While previous studies experimentally showed the similarities between model merging and ensembling, theoretical evidence and evaluation metrics remain lacking. To address this gap, we introduce Merging-ensembling loss (M-loss), a novel evaluation metric that quantifies the compatibility of merging source models using very limited unlabeled data. By measuring the discrepancy between parameter averaging and model ensembling at layer and node levels, M-loss facilitates more effective merging strategies. Specifically, M-loss serves both as a quantitative criterion of the theoretical feasibility of model merging, and a guide for parameter significance in model pruning. Our theoretical analysis and empirical evaluations demonstrate that incorporating M-loss into the merging process significantly improves the alignment between merged models and model ensembling, providing a scalable and efficient framework for accurate model consolidation.
Multi-task reinforcement learning (MTRL) seeks to learn a unified policy for diverse tasks, but often suffers from gradient conflicts across tasks. Existing masking-based methods attempt to mitigate such conflicts by assigning task-specific parameter masks. However, our empirical study shows that coarse-grained binary masks have the problem of over-suppressing key conflicting parameters, hindering knowledge sharing across tasks. Moreover, different tasks exhibit varying conflict levels, yet existing methods use a one-size-fits-all fixed sparsity strategy to keep training stability and performance, which proves inadequate. These limitations hinder the model’s generalization and learning efficiency. To address these issues, we propose SoCo-DT, a Soft Conflict-resolution method based by parameter importance. By leveraging Fisher information, mask values are dynamically adjusted to retain important parameters while suppressing conflicting ones. In addition, we introduce a dynamic sparsity adjustment strategy based on the Interquartile Range (IQR), which constructs task-specific thresholding schemes using the distribution of conflict and harmony scores during training. To enable adaptive sparsity evolution throughout training, we further incorporate an asymmetric cosine annealing schedule to continuously update the threshold. Experimental results on the Meta-World benchmark show that SoCo-DT outperforms the state-of-the-art method by 7.6% on MT50 and by 10.5% on the suboptimal dataset, demonstrating its effectiveness in mitigating gradient conflicts and improving overall multi-task performance.
Interactive preference elicitation (IPE) aims to substantially reduce human effort while acquiring human preferences in wide personalization systems. Dueling bandit (DB) algorithms enable optimal decision-making in IPE building on pairwise comparisons. However, they remain inefficient when human feedback is sparse. Existing methods address sparsity by heavily relying on parametric reward models, whose rigid assumptions are vulnerable to misspecification. In contrast, we explore an alternative perspective based on feedback augmentation, and introduce critical improvements to the model-free DB framework. Specifically, we introduce augmented confidence bounds to integrate augmented human feedback under generalized concentration properties, and analyze the multi-factored performance trade-off via regret analysis. Our prototype algorithm achieves competitive performance across several IPE benchmarks, including recommendation, multi-objective optimization, and response optimization for large language models, demonstrating the potential of our approach for provably efficient IPE in broader applications.
Multi-view clustering (MVC) aims to uncover the latent structure of multi-view data by learning view-common and view-specific information. Although recent studies have explored hyperbolic representations for better tackling the representation gap between different views, they focus primarily on instance-level alignment and neglect global semantic consistency, rendering them vulnerable to view-specific information (e.g., noise and cross-view discrepancies). To this end, this paper proposes a novel Wasserstein-Aligned Hyperbolic (WAH) framework for multi-view clustering. Specifically, our method exploits a view-specific hyperbolic encoder for each view to embed features into the Lorentz manifold for hierarchical semantic modeling. Whereafter, a global semantic loss based on the hyperbolic sliced-Wasserstein distance is introduced to align manifold distributions across views. This is followed by soft cluster assignments to encourage cross-view semantic consistency. Extensive experiments on multiple benchmarking datasets show that our method can achieve SOTA clustering performance.
ST-TPP: Learning Semi-Transductive Temporal Point Processes with Gromov-Wasserstein Barycentric Regularization
PDF ↗The generative mechanisms behind real-world event sequences are often heterogeneous, leading to data that possesses inherent clustering structures. However, most existing temporal point processes (TPPs) treat different event sequences independently, without leveraging the clustering structures when predicting events. In this study, we design and learn a novel semi-transductive temporal point process (ST-TPP), which explicitly improves prediction performance by co-training sequence clusters. In particular, given a set of event sequences, our method learns a neural TPP together with cluster centers of the sequences. Besides maximizing the likelihood of the event sequences, we leverage a data-based kernel matrix and prior knowledge to regularize the sequence embeddings, leading to a Gromov-Wasserstein barycentric (GWB) regularizer. Based on the optimal transport plans associated with the GWB regularizer, we derive the cluster centers by the push-forward of the sequence embeddings. When a new sequence comes, the learned model first assigns a cluster center to the sequence and then jointly encodes the sequence and the cluster center to predict future events, leading to a semi-transductive prediction scheme. Experiments demonstrate that ST-TPP achieves competitive sequence clustering results and strong prediction performance.
Self-Supervised Learning Based on Transformed Image Reconstruction for Equivariance-Coherent Feature Representation
PDF ↗Self-supervised learning (SSL) methods have achieved remarkable success in learning image representations allowing invariances in them — but therefore discarding transformation information that some computer vision tasks actually require. While recent approaches attempt to address this limitation by learning equivariant features using linear operators in feature space, they impose restrictive assumptions that constrain flexibility and generalization. We introduce a weaker definition for the transformation relation between image and feature space denoted as equivariance-coherence. We propose a novel SSL auxillary task that learns equivariance-coherent representations through intermediate transformation reconstruction, which can be integrated with existing joint embedding SSL methods. Our key idea is to reconstruct images at intermediate points along transformation paths, e.g. when training on 30° rotations, we reconstruct the 10° and 20° rotation states. Reconstructing intermediate states requires the transformation information used in augmentations, rather than suppressing it, and therefore fosters features containing the augmented transformation information. Our method decomposes feature vectors into invariant and equivariant parts, training them with standard SSL losses and reconstruction losses, respectively. We demonstrate substantial improvements on synthetic equivariance benchmarks while maintaining competitive performance on downstream tasks requiring invariant representations. The approach seamlessly integrates with existing SSL methods (iBOT, DINOv2) and consistently enhances performance across diverse tasks, including segmentation, detection, depth estimation, and video dense prediction. Our framework provides a practical way for augmenting SSL methods with equivariant capabilities while preserving invariant performance.
Marginal Fisher Analysis (MFA) is a classical dimensionality reduction (DR) method that leverages dual graphs to capture intra-class compactness and inter-class separability. However, MFA’s reliance on high-quality labels limits its practical application. For another, existing unsupervised DR methods neglect data’s local manifold relationship, resulting in poor discriminativeness. To address these limitations, we propose a novel DR method named Discriminative Graph Embedding Framework (DGEF) via Label-Free Marginal Fisher Analysis. Our approach uses the adjacency matrix and cluster indicator matrix derived from centerless K-Means to construct intrinsic graph and penalty graph, which preserve the local manifold structure of the data. Additionally, we have derived the convertible relationship between centerless K-Means and Manifold learning and unified them within a graph embedding framework. By adopting the intrinsic graph and penalty graph, our DGEF avoids centroid initialization and ensures robustness and discriminativeness. This method achieves dimensionality reduction adaptively without relying on labeled data. Extensive experiments on benchmark datasets show that our approach outperforms conventional methods in clustering performance.
Mixture-of-Experts (MoE) models enable scalable performance by activating large parameter sets sparsely, minimizing computational overhead. To mitigate the prohibitive cost of training MoEs from scratch, recent work employs upcycling, reusing a single pre-trained dense model by replicating its feed-forward network (FFN) layers into experts. However, this limits expert diversity, as all experts originate from a single pre-trained dense model. This paper addresses this limitation by constructing powerful MoE models using experts sourced from multiple identically-architected but disparate pre-trained models (e.g., Qwen2.5-Coder and Qwen2). A key challenge lies in the fact that these source models occupy disparate, dissonant regions of the parameter space, making direct upcycling prone to severe performance degradation. To overcome this, we propose Symphony-MoE, a novel two-stage framework designed to harmonize these models into a single, coherent expert mixture. First, we establish this harmony in a training-free manner: we construct a shared backbone via a layer-aware fusion strategy and, crucially, alleviate parameter misalignment among experts using activation-based functional alignment. Subsequently, a stage of post-training coordinates the entire architecture. Experiments demonstrate that our method successfully integrates experts from heterogeneous sources, achieving an MoE model that significantly surpasses baselines in multi-domain tasks and out-of-distribution generalization.
Label Distribution Learning (LDL) is an effective machine learning paradigm for addressing label ambiguity, where each sample is annotated with a distribution that conveys rich semantic information. However, during the actual annotation process of label distributions, annotators often exhibit divergent labeling preferences for the same sample. Most existing LDL methods overlook this heterogeneity, assuming that the observed label distribution originates from a single labeling pattern. Such an assumption limits their capacity to manage inter-annotator disagreement and constrains the generalization of the resulting models. To address this issue, we propose, for the first time, a Dirichlet process mixture model (DPMM)-based framework for LDL. This framework leverages nonparametric Bayesian methods to adaptively uncover diverse latent labeling patterns from the data and to accurately model annotator heterogeneity. Specifically, the ground-truth label distribution of each sample is modeled as a weighted mixture of multiple latent components, where a feature-conditioned gating mechanism adaptively controls the contribution of each component. Experimental results demonstrate that the proposed model consistently achieves competitive performance on several widely-used benchmark datasets.
Offline Meta-Reinforcement Learning with Flow-Based Task Inference and Adaptive Correction of Feature Overgeneralization
PDF ↗Offline meta-reinforcement learning (OMRL) combines the strengths of learning from diverse datasets in offline RL with the adaptability to new tasks of meta-RL, promising safe and efficient knowledge acquisition by RL agents. However, OMRL still suffers extrapolation errors due to out-of-distribution (OOD) actions, compromised by broad task distributions and Markov Decision Process (MDP) ambiguity in meta-RL setups. Existing research indicates that the generalization of the Q network affects the extrapolation error in offline RL. This paper investigates this relationship by decomposing the Q value into feature and weight components, observing that while decomposition enhances adaptability and convergence in the case of high-quality data, it often leads to policy degeneration or collapse in complex tasks. We observe that decomposed Q values introduce a large estimation bias when the feature encounters OOD samples, a phenomenon we term "feature overgeneralization''. To address this issue, we propose FLORA, which identifies OOD samples by modeling feature distributions and estimating their uncertainties. FLORA integrates a return feedback mechanism to adaptively adjust feature components. Furthermore, to learn precise task representations, FLORA explicitly models the complex task distribution using a chain of invertible transformations. We theoretically and empirically demonstrate that FLORA achieves rapid adaptation and meta-policy improvement compared to baselines across various environments.
RatioSketch: Towards More Accurate Frequency Estimation in Data Streams via a Lightweight Neural Network
PDF ↗Sketch-based solutions are widely used to estimate item frequencies in infinite data streams.Traditional hand-crafted sketches face the bottleneck of further eliminating errors because they cannot fully utilize the data stream distribution.Although recent neural sketches represented by MetaSketch and LegoSketch have improved generalization capabilities, they face bottlenecks such as high computational overhead and parameter sensitivity.Meanwhile, they ignore load information, fail to fully utilize the local information in hand-crafted sketches, and do not focus on the frequent items that are usually more important in data streams.In this paper, we propose RatioSketch, a novel lightweight neural network correction framework that synergizes the advantages of hand-crafted sketches and neural sketches in a ``micro-correction'' paradigm.The key idea is to retain the efficient underlying data structure of the hand-crafted sketch and to build a neural correction layer in its output space. We select multiple representative hand-crafted sketches as use cases to study the correction performance of RatioSketch on them.Extensive experimental evaluations on several real-world datasets show that RatioSketch-corrected sketches achieve consistently higher estimation accuracy than their uncorrected counterparts, as well as outperforming neural baselines such as MetaSketch and LegoSketch under identical memory budgets.
Spatially multimodal omics technologies provide unprecedented opportunities to address cellular heterogeneity within tissue contexts. However, learning robust and informative latent representations from such complex data remains a significant challenge. Existing graph-based methods often rely on static connections or indirect optimization objectives, which can constrain the discriminability and diversity of the learned representations, particularly in the presence of sequencing noise and unknown biological priors. To overcome these limitations, we propose Robust Integrative Analysis of Multi-omics Datasets via Nuclear-norm Maximization (RIA) to adaptively integrate multimodal features and spatial information through a new graph-based architecture. At the core of RIA is the introduction of the batch nuclear norm maximization (bnm) loss, marking the first application of bnm within the multi-omics domain. By maximizing the nuclear norm of the batch assignment matrix derived from the latent space, RIA simultaneously enhances the discriminability and diversity of the learned embeddings. This objective is synergistically combined with a dynamic prototype contrastive learning strategy and a graph stability loss, ensuring comprehensive and robust optimization.Ultimately, RIA produces a structured, information-rich latent space that enables more reliable downstream analyses, including cell type identification, spatial domain discovery, and microenvironment characterization.
Multi-graph multi-label learning (MGML) represents each object as a bag-of-graphs with multiple labels, but demands large-scale labeled data whose acquisition is often difficult and costly. Self-supervised contrastive learning (SCL) mitigates label dependence by leveraging data augmentation to construct discriminative pretext tasks, proving effective for multi-instance learning. However, when applied to MGML, SCL faces two key challenges: (1) it distinguishes individual instances by their differences, whereas MGML requires modeling label correlations; (2) it assumes semantic invariance under augmentation, but structural perturbations in MGML alter label semantics. To tackle these challenges, we propose a self-suPervised contrastive rE-learning framework for mulTi-grAph multi-labeL classification (PETAL). Specifically, to model label correlations, we first define a unified label space to learn label prototypes and align features with them, yielding prototype-aligned representations. We then design a multi-granularity contrastive loss over these representations, which captures label dependencies by contrasting at the bag level, graph level, and bag-graph level. Moreover, to ensure semantic invariance, we develop a contrastive re-learning strategy based on prototype-aligned representations to generate augmentation-free positive samples. This guarantees consistent multi-label distributions without structural perturbations. Experiments on six datasets demonstrate that PETAL achieves an average improvement of 4.12% over state-of-the-art self-supervised and supervised baselines.