Diffusion models have gained prominence as powerful generative tools for solving inverse problems due to their ability to model complex data distributions. However, existing methods typically rely on complete knowledge of the forward observation process to compute gradients for guided sampling, limiting their applicability in scenarios where such information is unavailable. In this work, we introduce *Constrained Particle Seeking (CPS)*, a novel gradient-free approach that leverages all candidate particle information to actively search for the optimal particle while incorporating constraints aligned with high-density regions of the unconditional prior. Unlike previous methods that passively select promising candidates, CPS reformulates the inverse problem as a constrained optimization task, enabling more flexible and efficient particle seeking. We demonstrate that CPS can effectively solve both image and scientific inverse problems, achieving results comparable to gradient-based methods while significantly outperforming gradient-free alternatives.
论文检索
输入标题、作者或关键词，从 100,903 篇学术成果中精准定位
Learning Dynamics as Feedback: An Adaptive Entropy Flow Dynamics Framework for Long-tailed Human Action Recognition
PDF ↗Deep human action recognition models trained on real-world data are often challenged by long-tailed distributions, where performance on rare classes is severely degraded. Current solutions typically apply static or heuristic interventions that are disconnected from the model's evolving internal state. To overcome this limitation, we reconceptualize long-tailed human action recognition as a closed-loop, self-regulating system, inspired by ecological theory. We further introduce an Adaptive Ecological Entropy Dynamics (AEED) framework, which is built upon three synergistic components. First, AEED perceives the learning state through entropy flow, providing a robust and directional signal of learning progress. Second, this signal drives an adaptation mechanism, which dynamically adjusts class-specific loss weights to allocate more learning resources to underperforming classes. Finally, AEED facilitates intelligent knowledge transfer via Confidence-Guided Symbiosis (CS-Mix). Extensive experiments demonstrate that AEED achieves state-of-the-art performance on challenging skeleton-based action recognition benchmarks, including NTU-60-LT and Kinetics-400-LT.
EnViT: Enhancing the Performance of Early-Exit Vision Transformers via Exit-Aware Structured Dropout-Enabled Self-Distillation
PDF ↗Vision Transformers (ViTs) have gained significant attention and widespread adoption due to their impressive performance in various computer vision tasks. However, in practice, their substantial computational overhead often leads to high inference latency and increased overheads when deployed on resource-constrained edge devices like smartphones, autonomous vehicles, and robots. To address these challenges, Early Exit (EE) has emerged as a promising approach for lightweight inference on edge devices. It accelerates inference and reduces computational overhead by adaptively producing predictions through early exits based on sample complexity. Existing EE methods typically suffer from substantial accuracy decreases in late exits while providing only marginal accuracy improvements to early exits. This paper presents EnViT, an exit-aware structured dropout-enabled self-distillation approach that enhances the performance of early exits without compromising late exits. EnViT leverages structured dropout to enable self-distillation, where the full model serves as the teacher and its own virtual sub-models generated by structured dropout as students. This mechanism effectively distills knowledge from the full model to early exits and avoids performance degradation in late exits by mitigating parameter conflicts across exits during training. Evaluation on five datasets shows that our EnViT achieves accuracy improvements ranging from 0.36% to 7.92% while maintaining competitive speed-up ratios of 1.72x to 2.23x.
Large Language Models (LLMs) exhibit pronounced memory-bound characteristics during inference due to High Bandwidth Memory (HBM) bandwidth constraints. In this paper, we propose an L2 Cache-oriented asynchronous KV Cache prefetching method to break through the memory bandwidth bottleneck in LLM inference through computation-load overlap. By strategically scheduling idle memory bandwidth during active computation windows, our method proactively prefetches required KV Cache into GPU L2 cache, enabling high-speed L2 cache hits for subsequent accesses and effectively hiding HBM access latency within computational cycles. Extensive experiments on NVIDIA H20 GPUs demonstrate that the proposed method achieves 2.15× improvement in attention kernel efficiency and up to 1.97× end-to-end throughput enhancement, surpassing state-of-the-art baseline FlashAttention-3. Notably, our solution maintains orthogonality to existing optimization techniques and can be integrated with current inference frameworks, providing a scalable latency-hiding solution for next-generation LLM inference engines.
Multi-view clustering (MVC), which aims to separate the multi-view data into distinct clusters in an unsupervised manner, is a fundamental yet challenging task. To enhance its applicability in real-world scenarios, this paper addresses a more challenging task: MVC under multi-source noises, including missing noise and observation noise. To this end, we propose a novel framework, Reliability-Aware Contrastive Deep Multi-View Clustering (RAC-DMVC), which constructs a reliability graph to guide robust representation learning under noisy environments. Specifically, to address observation noise, we introduce a cross-view reconstruction to enhances robustness at the data level, and a reliability-aware noise contrastive learning to mitigates bias in positive and negative pairs selection caused by noisy representations. To handle missing noise, we design a dual-attention imputation to capture shared information across views while preserving view-specific features. In addition, a self-supervised cluster distillation module further refines the learned representations and improves the clustering performance. Extensive experiments on five benchmark datasets demonstrate that RAC-DMVC outperforms SOTA methods on multiple evaluation metrics and maintains excellent performance under varying ratios of noise.
Depth-wise pruning accelerates LLM inference in resource-constrained scenarios but suffers from performance degradation due to indiscriminate removal of entire Transformer layers. This paper reveals ``Patch-Like'' redundancy across layers via correlation analysis of the outputs of different layers in reproducing kernel Hilbert space, demonstrating consecutive layers exhibit high functional similarity. Building on this observation, this paper proposes Sliding-Window Merging (SWM) - a dynamic compression method that selects consecutive layers from top to bottom using a pre-defined similarity threshold, and compacts patch-redundant layers through a parameter consolidation, thereby simplifying the model structure while maintaining its performance. Extensive experiments on LLMs with various architectures and different parameter scales show that our method outperforms existing pruning techniques in both zero-shot inference performance and retraining recovery quality after pruning. In particular, in the experiment with 35\% pruning on the Vicuna-7B model, our method achieved a 1.654\% improvement in average performance on zero-shot tasks compared to the existing method. Moreover, we further reveal the potential of combining depth pruning with width pruning to enhance the pruning effect.
Real-world visual data rarely presents as isolated, static instances. Instead, it often evolves gradually over time through variations in pose, lighting, object state, or scene context. However, conventional classifiers are typically trained under the assumption of temporal independence, limiting their ability to capture such dynamics. We propose a simple yet effective framework that equips standard feedforward classifiers with temporal reasoning, all without modifying model architectures or introducing recurrent modules. At the heart of our approach is a novel Support-Exemplar-Query (SEQ) learning paradigm, which structures training data into temporally coherent trajectories. These trajectories enable the model to learn class-specific temporal prototypes and align prediction sequences via a differentiable soft-DTW loss. A multi-term objective further promotes semantic consistency and temporal smoothness. By interpreting input sequences as evolving feature trajectories, our method introduces a strong temporal inductive bias through loss design alone. This proves highly effective in both static and temporal tasks: it enhances performance on fine-grained and ultra-fine-grained image classification, and delivers precise, temporally consistent predictions in video anomaly detection. Despite its simplicity, our approach bridges static and temporal learning in a modular and data-efficient manner, requiring only a simple classifier on top of pre-extracted features.
Prefix adders are widely used in compute-intensive applications for their high speed. However, designing optimized prefix adders is challenging due to strict design rules and an exponentially large design space. We introduce PrefixGPT, a generative pre-trained Transformer (GPT) that directly generates optimized prefix adders from scratch. Our approach represents an adder's topology as a two-dimensional coordinate sequence and applies a legality mask during generation, ensuring every design is valid by construction. PrefixGPT features a customized decoder-only Transformer architecture. The model is first pre-trained on a corpus of randomly synthesized valid prefix adders to learn design rules and then fine-tuned to navigate the design space for optimized design quality. Compared with existing works, PrefixGPT not only finds a new optimal design with a 7.7% improved area-delay product (ADP) but exhibits superior exploration quality, lowering the average ADP by up to 79.1%. This demonstrates the potential of GPT-style models to first master complex hardware design principles and then apply them for more efficient design optimization.
Communication-efficient Multi-Agent Reinforcement Learning with Spatiotemporal Information Hub
PDF ↗Centralized training with decentralized execution (CTDE) is a framework for MARL with wide applications. In the CTDE paradigm, agents leverage global state information during training to mitigate the non-stationarity of the MARL environment, but must rely solely on partial observations during execution. Recent work has highlighted the growing importance of inter-agent communication for more effective learning and coordination. However, most existing methods overlook the fact that real-world communication channels are often bandwidth-constrained and imperfectly reliable. Toward more communication-efficient and robust MARL, we extend the conventional CTDE framework with an information hub. The hub collects local observations from the agents to restore the global state, which is then delivered to the agents on demand. To this end, technical mechanisms are designed to enable effective global reconstruction with incomplete observations, as well as agent-specific attention to the reconstructed global information. Experiments on multiple cooperative MARL benchmarks demonstrate that our method achieves state-of-the-art performance compared to popular MARL algorithms while substantially reducing communication overhead and exhibiting strong robustness under imperfect communication channels.
TimeMosaic: Temporal Heterogeneity Guided Time Series Forecasting via Adaptive Granularity Patch and Segment-wise Decoding
PDF ↗Multivariate time series forecasting is essential in domains such as finance, transportation, climate, and energy. However, existing patch-based methods typically adopt fixed-length segmentation, overlooking the heterogeneity of local temporal dynamics and the decoding heterogeneity of forecasting. Such designs lose details in information-dense regions, introduce redundancy in stable segments, and fail to capture the distinct complexities of short-term and long-term horizons. We propose TimeMosaic, a forecasting framework that aims to address temporal heterogeneity. TimeMosaic employs adaptive patch embedding to dynamically adjust granularity according to local information density, balancing motif reuse with structural clarity while preserving temporal continuity. In addition, it introduces segment-wise decoding that treats each prediction horizon as a related subtask and adapts to horizon-specific difficulty and information requirements, rather than applying a single uniform decoder. Extensive evaluations on benchmark datasets demonstrate that TimeMosaic delivers consistent improvements over existing methods, and our model trained on the large-scale corpus with 321 billion observations achieves performance competitive with state-of-the-art TSFMs.
Diffusion models have recently shown promise in time series forecasting, particularly for probabilistic predictions. However, they often fail to achieve state-of-the-art point estimation performance compared to regression-based methods. This limitation stems from difficulties in providing sufficient contextual bias to track distribution shifts and in balancing output diversity with the stability and precision required for point forecasts. Existing diffusion-based approaches mainly focus on full-distribution modeling under probabilistic frameworks, often with likelihood maximization objectives, while paying little attention to dedicated strategies for high-accuracy point estimation. Moreover, other existing point prediction diffusion methods frequently rely on pre-trained or jointly trained mature models for contextual bias, sacrificing the generative flexibility of diffusion models. To address these challenges, we propose SimDiff, a single-stage, end-to-end framework. SimDiff employs a single unified Transformer network carefully tailored to serve as both denoiser and predictor, eliminating the need for external pre-trained or jointly trained regressors. It achieves state-of-the-art point estimation performance by leveraging intrinsic output diversity and improving mean squared error accuracy through multiple inference ensembling. Key innovations, including normalization independence and the median-of-means estimator, further enhance adaptability and stability. Extensive experiments demonstrate that SimDiff significantly outperforms existing methods in time series point forecasting.
Continual learning constrains models to learn new tasks over time without forgetting what they have already learned. A key challenge in this setting is catastrophic forgetting, where learning new information causes the model to lose its performance on previous tasks. Recently, explainable AI has been proposed as a promising way to better understand and reduce forgetting. In particular, self-explainable models are useful because they generate explanations during prediction, which can help preserve knowledge. However, most existing explainable approaches use post-hoc explanations or require additional memory for each new task, resulting in limited scalability. In this work, we introduce CIP-Net, an exemplar-free self-explainable prototype-based model designed for continual learning. CIP-Net avoids storing past examples and maintains a simple architecture, while still providing useful explanations and strong performance. We demonstrate that CIP-Net achieves state-of-the-art performances compared to previous exemplar-free and self-explainable methods in both task- and class-incremental settings, while bearing significantly lower memory-related overhead. This makes it a practical and interpretable solution for continual learning.
Diffusion models (DMs) memorize training images and can reproduce near-duplicates during generation. Current detection methods identify verbatim memorization but fail to capture two critical aspects: quantifying partial memorization occurring in small image regions, and memorization patterns beyond specific prompt-image pairs. To address these limitations, we propose Foreground Background Memorization (FB-Mem), a novel segmentation-based metric that classifies and quantifies memorized regions within generated images. Our method reveals that memorization is more pervasive than previously understood: (1) individual generations from single prompts may be linked to clusters of similar training images, revealing complex memorization patterns that extend beyond one-to-one correspondences; and (2) existing model-level mitigation methods, such as neuron deactivation and pruning, fail to eliminate local memorization, which persists particularly in foreground regions. Our work establishes an effective framework for measuring memorization in diffusion models, demonstrates the inadequacy of current mitigation approaches, and proposes a stronger mitigation method using a clustering approach.
The learning dynamics of modern neural networks remain an open problem in deep learning. The Neural Tangent Kernel (NTK) offers an elegant description of training dynamics in the infinite‑width limit, yet its classical formulation assumes a static data set. Modern model training practice departs from this strong assumption through the use of on‑the‑fly data augmentations (e.g. additive noise). In this work, we conduct an NTK-driven analysis of how data transformations affect a neural net's evolution in the function space. Our theoretical contributions characterize how repeated Gaussian perturbations from NTK-derived covariances can steer neural-net optimizations toward user‑specified behavior. These theoretical insights are empirically validated by controlled experiments. Taken together, our results lay the foundation for a promising future research direction that transforms the NTK from a descriptive to a prescriptive tool, enabling control of neural net training trajectories and behavior of inference generalization with grounded interventions.
Random Amalgamation of Adapters for Flatter Loss Landscapes: Towards Class-Incremental Learning with Better Stability
PDF ↗Class-incremental learning (CIL) enables models to continuously learn from streaming data while mitigating catastrophic forgetting of prior knowledge. Our research reveals that the CIL performance of pre-trained models (PTMs) varies significantly across different datasets, a phenomenon underexplored in existing studies. Through visualization, we observe that flatter loss landscapes correlate with superior CIL performance. This insight motivates us to enhance PTMs' CIL capability by promoting loss landscapes' flatness. Initially, we propose independently optimizing multiple adapter branches to equip PTMs with diverse learnable parameters, thereby improving stability during parameter updates. However, given computational and memory constraints, the number of adapters a PTM can accommodate is limited. To address this, we introduce a training strategy with randomized adapter amalgamation (RAA), compelling the model to maintain low loss across a broader and more continuous parameter space, significantly enhancing flatness. Furthermore, we refine existing sharpness-aware minimization techniques to further optimize the loss landscapes. Our extensive experiments and visualization results validate the efficacy of the method, resulting in the state-of-the-art (SOTA) performance.
Split inference (SI) enables users to access deep learning (DL) services without directly transmitting raw data. However, recent studies reveal that data reconstruction attacks (DRAs) can recover the original inputs from the smashed data sent from the client to the server, leading to significant privacy leakage. While various defenses have been proposed, they often result in substantial utility degradation, particularly when the client-side model is shallow. We identify a key cause of this trade-off: existing defenses apply excessive perturbation to redundant information in the smashed data. To address this issue in computer vision tasks, we propose InfoDecom, a defense framework that first decomposes and removes redundant information and then injects noise calibrated to provide theoretically guaranteed privacy. Experiments demonstrate that InfoDecom achieves a superior utility-privacy trade-off compared to existing baselines.
Efficiently Seeking Flat Minima for Better Generalization in Fine-Tuning Large Language Models and Beyond
PDF ↗Little research explores the correlation between the expressive ability and generalization ability of the low-rank adaptation (LoRA). Sharpness-Aware Minimization (SAM) improves model generalization for both Convolutional Neural Networks (CNNs) and Transformers by encouraging convergence to locally flat minima. However, the connection between sharpness and generalization has not been fully explored for LoRA due to the lack of tools to either empirically seek flat minima or develop theoretical methods. In this work, we propose Flat Minima LoRA (FMLoRA) and its efficient version i.e., EFMLoRA, to seek flat minima for LoRA. Concretely, we theoretically demonstrate that perturbations in the full parameter space can be transferred to the low-rank subspace. This approach eliminates the potential interference introduced by perturbations across multiple matrices in the low-rank subspace. Our extensive experiments on large language models and vision-language models demonstrate that EFMLoRA achieves optimization efficiency comparable to that of LoRA while simultaneously attaining comparable or even better performance. For example, on the GLUE dataset with RoBERTa-large, EFMLoRA outperforms LoRA and full fine-tuning by 1.0% and 0.5% on average, respectively. On vision-language models e.g., Qwen-VL-Chat, there are performance improvements of 1.5% and 1.0% on the SQA and VizWiz datasets, respectively. These empirical results also verify that the generalization of LoRA is closely related to sharpness, which is omitted by previous methods.
FedSkeleton: Secure Multi-Party Graph Skeleton Construction for Privacy-Preserving Federated Time-Series Forecasting
PDF ↗In real-world time-series modelling, graph structures are widely adopted because they explicitly encode node topology and capture complex network dynamics. In practice, however, a complete graph is often partitioned across multiple parties; each party can access only its local sub-graph and, owing to privacy regulations, cannot share topology or data, creating pervasive data silos. Federated Graph Learning (FGL) offers a privacy-preserving collaborative-learning paradigm, yet current methods still face two key challenges: (1) the graph topology itself contains sensitive structural information, which can lead to privacy leakage if directly shared during FGL; (2) cross-party edges are crucial for accurate modeling, yet exploiting them without compromising privacy remains a significant challenge. To overcome these obstacles, we propose FedSkeleton, a privacy-preserving framework for time-series prediction that comprises a Skeleton Construction Module and a Dual-stream Forecasting Module, enabling global dependency capture without revealing the topology. Extensive experiments show that FedSkeleton consistently outperforms existing baselines and even surpasses models trained in a centralized setting with full-graph access in certain cases. In addition, we conduct comprehensive security analysis, communication-cost evaluation and scalability experiments, demonstrating that FedSkeleton effectively resists common attacks, keeps communication overhead manageable, and remains robust with respect to key hyper-parameters and the number of participating parties.
Understanding how localized changes in one variable affect others in multivariate time series is essential for diagnostics and decision-making in complex systems. Existing models often fail to capture realistic inter-feature dynamics when simulating "what-if" scenarios, leading to inaccurate or uncorrelated reconstructions. We propose CFORVAE, a variational autoencoder framework that explicitly addresses this limitation by combining temporal decomposition with frequency-domain feature correlation modeling. Our architecture uses a dual-path encoding of trend and seasonal components, each projected into attention-pooled latent spaces, and applies Fourier Neural Operators (FNO) to capture cross-feature dependencies in the spectral domain. This decomposition-correlation design enables component-specific latent manipulation and ensures that local modifications propagate realistically across correlated variables. Through extensive experiments, we show that CFORVAE outperforms state-of-the-art baselines in preserving temporal and feature-level dependencies, especially under adjustment-based reconstructions, making it a powerful tool for interpretable "what-if" analysis and diagnostics.
Deep learning models often achieve high accuracy but lack interpretability, making them unsuitable for critical applications such as medical diagnosis, biomolecule design, criminal justice, etc. The Sparse High-order Interaction Model (SHIM) addresses this limitation by providing both transparency and predictive reliability. However, real-world data often contain outliers, which can distort model performance. To overcome this, we propose Huberized-SHIM, an extension of SHIM that integrates Huber loss-based robust regression to mitigate the impact of outliers. We introduce a homotopy-based exact regularization path algorithm and a novel tree-pruning criterion to efficiently manage interaction complexity. Additionally, we incorporate the conformal prediction framework to enhance statistical reliability. Empirical evaluations on synthetic and real-world datasets demonstrate the superior robustness and accuracy of Huberized-SHIM in high-stakes decision-making contexts.