Self-supervised representation learning for time series has largely bifurcated into contrastive and generative paradigms. However, both approaches can face challenges in capturing underlying physical dynamics: contrastive methods risk altering temporal semantics through heuristic spatial augmentations, while generative masked modeling tends to overfit high-frequency sensor noise rather than abstracting stable evolution laws. In this work, we posit that the semantic identity of a time series is defined not only by static patterns, but also by its invariant laws of transition. To operationalize this, we introduce K-VIC, a novel framework that embeds the Koopman Operator into a Variance-Covariance regularization scheme. By linearizing non-linear temporal dynamics in a high-dimensional latent space, K-VIC acts as a physics-inspired spectral filter that effectively preserves persistent dynamic modes while discarding non-systematic fluctuations. Furthermore, instead of relying on augmented negative pairs or static spatial invariance, we propose predictive and cyclic consistency paired with informational constraints to mitigate representation collapse and encourage orthogonal feature disentanglement. Empirical evaluations across five benchmarks demonstrate that K-VIC achieves state-of-the-art (SotA) accuracy without dataset-specific hyperparameter tuning. Notably, it exhibits strong robustness in noise-heavy industrial scenarios (yielding an over 20 https://github.com/Conrrile/K-VIC .
We study stochastic bandits with ordered actions, unimodal rewards, and resource constraints, motivated by treatment selection problems where intervention intensity improves outcomes up to a peak while incurring increasingly higher costs. Unlike existing constrained bandit approaches, whose regret bounds typically scale with the number of actions because feasibility and optimality must be explored across an unstructured action set, our setting incorporates two key forms of structure: unimodality of rewards, which reduces optimality learning to local exploration around the empirical peak, and monotonicity of costs, which implies that the feasible region is a contiguous prefix that can be identified through a single threshold. We propose F-OSUB, an algorithm that interleaves feasibility identification with unimodal leader–neighbor exploration and show that it achieves logarithmic regret and logarithmic budget violation with high probability, with constants depending only on local reward gaps rather than the total number of actions. These results demonstrate that exploiting structural properties enables substantially more efficient and safer learning in resource-constrained decision problems such as treatment intensity selection.
Diffusion and flow matching are a class of generative models that generate new samples by solving ordinary or stochastic differential equations with a learned score/velocity field. Interestingly, the optimal velocity field admits a closed-form formula that can be computed for finite datasets. Generating samples following the optimal velocity field can only reproduce samples from the training set, i.e., memorize the training set. Neural networks, trained to match the velocity field, introduce inductive bias that can partially mitigate the issue. However, these models can still exhibit memorization, unlike the benign overfitting observed in discriminative tasks. In this paper, we propose a bootstrap aggregating (bagging) method for flow matching to reduce memorization. By conceptually averaging over resampled training subsets, our approach effectively reduces memorization. We derive a closed-form bagging formulation compatible with exact flow matching, enabling efficient implementation. Experiments confirm reduced memorization and better generalization without architectural changes or auxiliary objectives.
The Centralized Training with Decentralized Execution (CTDE) paradigm has become increasingly popular in multi-agent reinforcement learning and is widely adopted in recent works. However, decentralized policies operate under partial observations and may achieve suboptimal performance compared to centralized policies, while naive centralized policies often struggle to scale with larger numbers of agents. To address these limitations, we introduce Centralized Permutation Equivariant (CPE) learning, a scalable centralized training and execution (CTE) framework that transforms standard CTDE algorithms by replacing decentralized execution with a fully centralized policy, while retaining their centralized training components. Our policy network is built upon a Global–Local Permutation Equivariant (GLPE) architecture, which is lightweight, computationally efficient, and agent-number-agnostic at the architectural level. Empirical results show that CPE can be seamlessly integrated with both value decomposition and actor–critic methods, consistently improving the performance of classical CTDE approaches across cooperative benchmarks such as MPE, SMAC, and RWARE, while matching state-of-the-art performance on RWARE. These results suggest that scalable centralized execution constitutes a promising alternative to decentralized policies in coordination-intensive MARL settings.
Theoretically, proper ensemble of multiple different models can improve the recommendation effectiveness. Unfortunately, most of the existing ensemble methods which depend on static aggregation strategies are not suitable for sequential recommendation, because they fail to distinguish the differences in predictive capabilities of base models among various users and contexts. In addition, most of the existing diversity measures, which are used in ensemble learning to strengthen the generalization performance, are designed for classification or regression tasks and not suitable for the ranking tasks in recommender systems. To tackle the above problems, we propose a Reinforcement Learning-based Adaptive Ensemble framework for sequential Recommendation (RLAERec), to effectively aggregate different base models for sequential recommendation. Specifically, we model the ensemble process as a Markov decision process (MDP) and design a dynamic weight allocation method according to constantly changing user behavior patterns and feedback. In addition, we propose an adaptive diversity measure for ranking ensemble learning, which can further enhance the ensemble performance. Extensive experiments on four real-world datasets demonstrate the effectiveness of the proposed model.
Low-dimensional tabular datasets often exhibit noise, heterogeneity, and limited expressive capacity, while common pipelines rely on fixed expansions or manual feature engineering. We introduce AugTab, a modular framework that makes learnable feature augmentation a first-class component for tabular learning. At its core is a Feature Augmentation Layer (FAL) that expands inputs via complementary branches for nonlinear projections, explicit cross-interactions, and gated recombinations, all trained jointly with the downstream backbone. To address low-dimensional tabular constraints, AugTab includes regularizers for compute budget, feature stability, and robustness to distribution shift, enabling efficient and resilient deployments. The layer is architecture-agnostic and integrates with MLPs. Across 15 low-dimensional benchmarks spanning diverse domains, AugTab improves performance, achieves the best average rank across classification tasks, and attains the top result on all regression tasks, outperforming 54 classification baselines and 16 regression baselines, respectively. These results position learnable feature augmentation as a practical foundation for robust low-dimensional tabular learning.
Vector quantization is a fundamental technique for reducing memory footprint and computational overhead in modern database systems, search engines, and large language models. However, existing vector quantization methods remain limited in simultaneously achieving high quantization accuracy and computational efficiency. In this paper, we propose Distribution-aware locally-adaptive Quantization (DalQ), which achieves superior accuracy with competitive efficiency through two key mechanisms. First, distribution-aware range adaptation leverages Gaussian-like embedding distributions via an adaptive clip factor, concentrating precision in high-density regions while tolerating controlled clipping of rare outliers. Second, decomposed vector refinement employs a two-stage geometric optimization strategy for near-optimal reconstruction: directional alignment via iterative code search and magnitude refinement via projection-based scaling. We prove the existence and uniqueness of the optimal clip factor and identify a reference clip factor enabling tuning-free deployment. Extensive evaluations demonstrate the substantial advantages of DalQ over state-of-the-art baselines. Against LVQ, DalQ achieves up to 96.6 × faster quantization and up to 3.2 × faster distance computation.
We show that grokking, the sudden generalization of overfit neural networks, coincides with a geometric phase transition in representation space. Using Archetypal Analysis (AA), we decompose residual-stream activations of single-layer Transformers trained on modular arithmetic into convex combinations of extreme points. Before grokking, activations are diffuse. At the transition, they reorganize into a simplex whose P vertices correspond one-to-one with output classes, and this co-emergence serves as a mechanistic signature of grokking. The simplex structure is then causally validated by direct intervention: swapping a sample’s AA mixture weights to those of a target class redirects the model’s prediction with 98– 100% accuracy, while k-means ( ∼6% ) and unstructured baselines fail. The decomposition requires no label supervision: overcomplete AA with behavior-based merging recovers exactly P groups with 100% label alignment across all seeds. Archetype permutations further reveal cyclic group structure, implementing y ↦ y+d P at 80– 85% with k=P and 100% with the behavior-merged k = 2P→P representation, connecting the simplex geometry to the Fourier mode theory of grokking.
Event prediction significantly affects public safety, healthcare, and resource optimization. Current methods learn representations of past events through temporal event graphs to predict whether a future event will occur. However, the effectiveness of these methods depends on rich event labels and low-noise data, which are often unattainable in real-world scenarios. In this paper, we propose a Semantics-preserving Temporal Adversarial Graph Contrastive Learning (Sead) method to address this issue. This is achieved by contrasting the clean graph with three auxiliary views (i.e., the semantics-preserving view, the adversarial view and the temporal perturbation view) to learn robust event representations while preserving the associations between events. Specifically, Sead first generates auxiliary views based on the semantic similarity between events and effective perturbations to the temporal event graph. Then, Sead introduces a semantics-aware event encoder that effectively captures the complete semantics of each event. Finally, Sead obtains robust event representations to predict future events. Experimental results demonstrate that our method outperforms current state-of-the-art approaches on real-world datasets. The code is available at https://github.com/hucheng-IIE/Sead .
While multimodal recommendation models have effectively integrated visual and textual information, their reliance on unique ID embeddings constitutes a fundamental performance bottleneck. Specifically, ID-based paradigms suffer from three limitations: (1) Information Isolation, where unique IDs prevent semantic information exchange among related items; (2) Cold-Start Vulnerability, as ID embeddings are difficult to optimize with sparse interactions; and (3) Storage Inefficiency, where parameter costs scale linearly with item quantity. To overcome these challenges, we propose MOTOR, a novel ID-free MultimOdal TOken Representation scheme. MOTOR replaces explicit item IDs with learnable, shared multimodal tokens, fundamentally transforming the recommender into an ID-free framework. Methodologically, we first employ product quantization to discretize raw multimodal features into compact token IDs. These tokens serve as implicit item features, which are then synthesized via a novel Token Cross Network (TCN) to capture high-order interaction patterns. This "discretize-and-interact" mechanism enables semantic sharing across items and significantly compresses the model size without introducing complex auxiliary losses. Extensive experiments across nine mainstream models demonstrate the significant performance improvement achieved by MOTOR. Further, MOTOR improves the capability of these models to recommend items in cold-start scenarios.
Training deep neural networks (DNNs) under high sparsity remains challenging, as model performance often degrades significantly once sparsity exceeds a critical threshold. Dynamic Sparse Training (DST) has emerged as an effective paradigm that dynamically evolves network connectivity during training, enabling sparse models to achieve better performance than static sparse training. Despite its empirical success, the optimization mechanisms underlying DST remain insufficiently understood. In this work, we investigate the training dynamics of DST and provide empirical evidence that connectivity evolution is associated with transient stochastic perturbations in the optimization trajectory. These perturbations are consistent with enhanced exploration of the loss landscape and may help sparse models move away from sharp or suboptimal regions. Motivated by this observation, we propose Trajectory-Regularized Stochasticity Injection (TRSI), a lightweight mechanism that enhances stochastic exploration while stabilizing training through trajectory regularization. TRSI introduces magnitude-aware stochastic perturbations to active parameters and incorporates trajectory-based regularization to guide optimization. Extensive experiments on CIFAR-100, TinyImageNet, and ImageNet demonstrate that TRSI consistently improves existing DST methods across multiple architectures and sparsity levels, with notable gains under extreme sparsity. These results highlight the importance of stochastic exploration in sparse training and provide a simple yet effective approach for improving dynamic sparse training.
We investigate the challenge of promoting diversity in offline reinforcement learning (RL), where agents must develop diverse strategies despite being trained on homogeneous datasets with limited behavioral variation. Existing offline RL approaches, including those leveraging expectation-maximization algorithms for unsupervised clustering, often struggle with either insufficient diversity or performance degradation in such settings. To overcome these limitations, we introduce a novel Unique Behavior objective function that can be directly computed to quantify the distinctiveness between agents, eliminating the need for additional estimators and reducing potential estimation errors. By maximizing uniqueness, our approach encourages agents to learn diverse behaviors effectively, even when the training data lacks variety. Extensive experiments on standard and diverse D4RL benchmarks, together with Atari evaluations, demonstrate that our method consistently achieves stronger quality-diversity trade-offs than DIVEOFF, CLUE, and SORL while maintaining competitive task performance across homogeneous and heterogeneous datasets.
The globalization of e-commerce platforms increases the need for effective Cross-Market Sequential Recommendation (CMSR) systems to alleviate data sparsity and the cold-start problem in emerging target markets. Current frameworks, operating under strict privacy regulations, utilize a “pre-train, then fine-tune“paradigm to transfer sequential patterns from data-rich source markets. While selectively transferring self-attention weights mitigates the negative transfer caused by market shifts, existing methods rely heavily on simple Multi-Layer Perceptron (MLP) adapters. These linear transformations often struggle to capture complex, non-linear discrepancies in item popularity and heterogeneous user preferences across different regions. In this paper, we present a comprehensive comparative analysis of advanced Parameter-Efficient Fine-Tuning (PEFT), reconstructive and generative adapters within the CMSR framework. We systematically evaluate Low-Rank Adaptation (LoRA), Encoder-Decoder, Denoising Autoencoder (DAE), and Variational Autoencoder (VAE) architectures against the baseline MLP. Extensive experiments on the real-world XMarket dataset demonstrate that reconstructive and generative paradigms significantly enhance transfer robustness. Notably, the DAE adapter achieves the highest average nDCG@10 performance among the evaluated adapters. The observed improvements are consistent with the hypothesis that reconstructive adapters provide more robust adaptation under target-market sparsity, with the DAE adapter achieving a substantial relative improvement in nDCG@10 over the MLP baseline. Our findings suggest that reconstructive adapters constitute a promising adaptation mechanism for mitigating market discrepancies in decentralized sequential recommendation.
Continual learning seeks to balance the stability-plasticity trade-off and address the catastrophic forgetting problem in deep neural networks. Architecture-based approaches tackle forgetting by assigning each task an isolated network subspace, which is either frozen for maximal stability or allowed small updates based on less fine-grained task information in recent gradient adjustment approaches. Notably, fine-grained task information has been shown to play a crucial role among other approaches, meaningfully contributing to guiding the training of new tasks. Inspired by this, we introduce fine-grained, neuron-level task information in this paper to gradient adjustment approaches to provide more accurate and effective training guidance. Building on the Hard Attention to the Task (HAT) architecture, we propose Fine-Grained Adaptive HAT (FG-AdaHAT), a general adaptive gradient adjustment framework for architecture-based continual learning, to which any neuron-wise importance measure and its scheduler can be applied. We propose several representative types of fine-grained importance measures for our framework constructed from training information and layer attribution methods. Experimental results show that FG-AdaHAT, using these importance measures, outperforms existing approaches on several benchmarks with long task sequences. Our code is available ( http://pengxiang-wang.com/projects/continual-learning-arena ).
While large language models (LLMs) excel at generating explanations in real-world recommendation scenarios, the constraints on response latency and deployment costs limit their practical use. As a solution, Knowledge Distillation has emerged as an efficient paradigm, providing a viable path to migrate the capabilities of LLMs into compact student models. However, transferring LLMs’ abilities for preference inference and long-form explanation generation into a lightweight student model still remains a challenge. First, there is substantial variation in the difficulty of analyzing user historical interaction records. Without considering the different levels of difficulty, the student model may learn rough and even inaccurate rationales. Second, the explanatory text itself is often noisy and tends to be generic. If the student model is trained directly to imitate the outputs of the teacher model, it tends to learn formulaic expressions rather than the underlying reasoning process. To address these issues, we present a rationale distillation framework that employs rationale generated by a teacher model as structured supervision. Particularly, we estimate difficulty scores from history-embedding clustering statistics and adopt an easy-to-hard curriculum that progressively introduces harder, i.e., more ambiguous instances. To reduce noise in teacher-generated rationales, we learn their token weights using a weight generator. The generator is trained with simple regularization and updated iteratively so that the weights remain effective as the student model evolves. Finally, we train the student model with a dual-weighted loss that includes both the instance difficulty and the token weights. Experiments on the MovieLens and Amazon Electronics datasets show that distilling Qwen2.5-32B into Qwen2.5-3B achieves satisfactory explanation as well as effective ranking performance for sequential recommendation with lower inference costs. ( ^1 The code is available at https://github.com/HduDBSI/RDistillRec .)
Offline Meta-Reinforcement Learning (OMRL) enables agents to generalize to unseen tasks using pre-collected static datasets. However, a critical challenge arises when these datasets are of mixed quality, originating from multi-modal behavior policies that blend expert demonstrations with sub-optimal explorations. While current approaches prioritize behavior-invariant purification, this strategy does not necessarily imply that the learned representation preserves the minimal sufficient statistics of task-defining transition and reward dynamics, resulting in an under-specified latent space that fails to resolve unseen task structures during inference. To overcome this, we propose Multi-modal Offline Decomposition for Expertise (MODE), a framework to address the challenge of learning optimal meta-policy from mixed-quality offline dataset. MODE consists of three components: (1) a Recurrent GM-VAE context model that incorporates Recurrent Neural Network (RNN) and Gaussian Mixture Model (GMM) on top of Variational Auto-encoder (VAE) architecture to robustly capture task representations from temporal features; (2) a GMM-based behavior decomposition process that distinguishes between varying quality modes within the dataset; and (3) a Lower Confidence Bound (LCB) selection mechanism that guides a hyper-policy to selectively combine the most promising Gaussian components based on value estimation and uncertainty. Experimental results demonstrate that our method consistently outperforms state-of-the-art baselines on continuous control benchmarks in MuJoCo, particularly in scenarios involving mixed-qualities. Our method not only achieves higher asymptotic returns, but also exhibits significantly lower variance, highlighting its robustness to task variations and data noise.
Many large-scale recommender systems adopt a two-stage architecture to balance computational efficiency and performance. Off-policy evaluation (OPE) of candidate generators in such two-stage recommender systems is challenging because action-set changes can violate the overlap assumption required by standard importance-weighting estimators. Existing proxy-based methods address this by aggregating actions into equivalence classes, but their reliance on heuristics can incur substantial approximation bias. In this work, we propose Latent-Proxy Alignment IPS (LPAIPS), which addresses this limitation by learning reward-aware latent proxies. LPAIPS learns a mapping from context–action pairs to discrete latent classes and promotes latent-space overlap through distribution alignment. As a theoretical contribution, we formalize sufficient conditions for latent overlap control and characterize the bias–variance trade-off in terms of mean squared error (MSE). Through experiments on both synthetic data and a real-world public dataset (Open Bandit Dataset), we show that LPAIPS achieves substantially lower MSE than IPS, IIPS, and Proxy IPS in severe support-mismatch settings, with up to a 9.2-fold reduction in MSE relative to Proxy IPS under severe support mismatch and up to a 74
Sequential recommendation predicts users’ next interactions by modeling the dynamic evolution of their interests from historical behavior sequences. Given that long-term and short-term interests are intrinsically intertwined within these sequences, existing methods struggle to capture this complexity accurately, facing two key challenges: (i) inadequate decoupling, which causes dominant long-term interests to overshadow localized short-term interests; and (ii) uniform processing strategies, which ignore the distinct modeling requirements of long-term and short-term interests, failing to distinguish informative short-term interests from incidental noise. To address these challenges, we propose a new framework, Hierarchical Wavelet Decoupling Framework with Adaptive Gating for Sequential Recommendation (HiWaveRec), which enables differentiated modeling of long-term and short-term interests. Specifically, HiWaveRec applies a multi-level Discrete Wavelet Transform (DWT) to decouple interaction sequences into low-frequency long-term interests and multi-scale high-frequency short-term interests while preserving temporal locality. It then employs a heterogeneous dual-branch architecture, where Low-Frequency Self-Attention (LFSA) captures global dependencies in long-term interests and Causal Detail Gating (CDG) selectively amplifies informative short-term interests while suppressing noise. Finally, the processed sub-bands are fused through Inverse DWT(IDWT) to reconstruct a unified sequence representation. Experiments on four benchmark datasets demonstrate that HiWaveRec achieves competitive or superior performance compared with strong baselines. The source code is available here ( https://github.com/coderxxw/HiWaveRec ).
Cooperative multi-agent reinforcement learning (MARL) aims to enable agents to achieve coordinated behaviors in complex environments. However, existing methods often struggle to strike a balance between rapid local adaptation and consistent long-term cooperation, resulting in poor coordination. To address this challenge, we propose Hierarchical Cognitive Learning (HCL), a framework inspired by hierarchical mechanisms of human decision-making. This framework models hierarchical cognitive processes by integrating short-term perceptual alignment for responsive coordination, long-term intent modeling for strategic abstraction, and cognitive policy generation that unifies both pathways to produce coherent joint behaviors. This design allows agents to dynamically balance immediate reactions and long-horizon planning under partial observability. We evaluate HCL on the Active Voltage Control (AVC) task, the StarCraft Multi-Agent Challenge (SMAC) benchmark, and the SMAC-Hard benchmark, covering both continuous and discrete action spaces. Experimental results demonstrate that HCL outperforms state-of-the-art methods, facilitating effective multi-agent coordination.
Reinforcement learning (RL) has emerged as a promising solution for addressing traffic signal control (TSC) challenges. While most RL-based TSC systems typically employ an online approach, facilitating frequent active interaction with the environment, learning such strategies in the real world is impractical due to safety and risk concerns. To tackle these challenges, this study introduces an innovative offline data-driven approach, called DataLight. DataLight employs effective state representations and a reward function by capturing vehicular speed information within the environment. It then segments roads to capture spatial information and further enhances the spatially segmented state representations with sequential modeling. The experimental results demonstrate the superior performance of DataLight compared to both state-of-the-art online and offline TSC methods. Additionally, DataLight exhibits robust learning capabilities concerning real-world deployment issues. Our source code is publicly available at https://github.com/LiangZhang1996/DataLight (The supplementary appendix for this paper is archived on Zenodo with DOI: https://doi.org/10.5281/zenodo.20635875 ).
