Multimodal large language models (LMMs) have demonstrated remarkable capabilities across diverse vision-language tasks, including image captioning, visual question answering, and text-image retrieval. However, their computational complexity and memory footprint, particularly in the key-value (KV) cache during inference, pose significant challenges for real-time deployment, especially on resource-constrained devices. In this paper, we propose Dynamic KV Cache Quantization, a novel quantization strategy tailored for multimodal LMMs. Our approach applies per-channel quantization to (K) and per-token quantization to (V), leveraging their respective statistical distributions to optimize precision allocation. Additionally, we introduce an adaptive token and channel recording mechanism that dynamically adjusts quantization parameters based on real-time distribution tracking, effectively mitigating the impact of outliers. To further enhance compression efficiency, we implement fine-grained grouping, which partitions KV tensors into localized subgroups, enabling more adaptive quantization. Experimental results on LLaVA-1.5 (7B/13B) and Qwen-VL across multiple multimodal benchmarks demonstrate that our method significantly outperforms existing KV-cache quantization approaches, achieving a superior trade-off between memory efficiency and model accuracy.
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
Decomposing and Composing: Towards Efficient Vision-Language Continual Learning via Rank-1 Expert Pool in a Single LoRA
PDF ↗Continual learning (CL) in vision-language models (VLMs) faces significant challenges in improving task adaptation and avoiding catastrophic forgetting. Existing methods usually have heavy inference burden or rely on external knowledge, while Low-Rank Adaptation (LoRA) has shown potential in reducing these issues by enabling parameter-efficient tuning. However, considering directly using LoRA to alleviate the catastrophic forgetting problem is non-trivial, we introduce a novel framework that restructures a single LoRA module as a decomposable Rank-1 Expert Pool. Our method learns to dynamically compose a sparse, task-specific update by selecting from this expert pool, guided by the semantics of the [CLS] token. In addition, we propose an Activation-Guided Orthogonal (AGO) loss that orthogonalizes critical parts of LoRA weights across tasks. This sparse composition and orthogonalization enable fewer parameter updates, resulting in domain-aware learning while minimizing inter-task interference and maintaining downstream task performance. Extensive experiments across multiple settings demonstrate state-of-the-art results in all metrics, surpassing zero-shot upper bounds in generalization. Notably, it reduces trainable parameters by 96.7% compared to the baseline method, eliminating reliance on external datasets or task-ID discriminators. The merged LoRAs retain less weights and incur no inference latency, making our method computationally lightweight.
Reliable prediction of train delays is essential for enhancing the robustness and efficiency of railway transportation systems. In this work, we reframe delay forecasting as a stochastic simulation task, modeling state-transition dynamics through imitation learning. We introduce Drift-Corrected Imitation Learning (DCIL), a novel self-supervised algorithm that extends DAgger by incorporating distance-based drift correction, thereby mitigating covariate shift during rollouts without requiring access to an external oracle or adversarial schemes. Our approach synthesizes the dynamical fidelity of event-driven models with the representational capacity of data-driven methods, enabling uncertainty-aware forecasting via Monte Carlo simulation. We evaluate DCIL using a comprehensive real-world dataset from \textsc{Infrabel}, the Belgian railway infrastructure manager, which encompasses over three million train movements. Our results, focused on predictions up to 30 minutes ahead, demonstrate superior predictive performance of DCIL over traditional regression models and behavioral cloning on deep learning architectures, highlighting its effectiveness in capturing the sequential and uncertain nature of delay propagation in large-scale networks.
FlorE: Integrating Full Lorentz Group and Directional Offsets for Effective Knowledge Graph Embedding
PDF ↗Knowledge Graph Embedding (KGE) aims to map entities and relationships into a continuous vector space to facilitate reasoning and downstream tasks. Although previous KGE methods based on Euclidean, complex spaces, or hyperbolic spaces have performed well, they still struggle to effectively model Z-Paradox relation patterns which account for a large proportion in each knowledge graph. To address this issue, we propose a novel KGE method **FlorE** which integrates full Lorentz Group and directional offset operation in hyperbolic space for KGE task. Specifically, we incorporates the full Lorentz Group to enable the same relation in knowledge graph (KG) to perform indefinite isometry, thus avoiding the overlapping of entities. Meanwhile, we implement directional offset operation via exponential mapping to transform the relations to the same Lorentz manifold of the entities, thus maintaining geometric consistency for the relations and entities in KG. By integrating these two techniques, FlorE can effectively model the Z-Paradox relation patterns and improve the representation learning ability for KGs. Experiments on the five benchmark datasets demonstrate that our method achieves state-of-the-art performance. For the Z-Paradox relation patterns, the improvement achieves **26.7%**, **15.6%**, **35.4%**, **33.7%**, and **31.5%** on FB15k-237, WN18RR, CoDEx-S, CoDEx-M and CoDEx-L, respectively.
Time series generation is essential for advancing data-driven modeling and decision-making across a wide range of domains. However, existing approaches primarily focus on global patterns, often failing to capture local key patterns such as abrupt changes or anomalies. These key patterns are crucial for interpretability and operational decision making, as they frequently represent intervention points with significant real-world impact. To bridge this gap, we propose Key Prototypes-Guided Diffusion (K-ProtoDiff) for time series generation , a new model that learns the global data distribution while preserving localized key patterns critical for temporal dynamics. In K-ProtoDiff, we first derive time series prototype representations through adaptive self-supervised learning. Then, a key prototype assignment module is used to extract prototype weights, forming key prototype-aware representations that serve as conditional guidance for generation. During sampling, to further enhance the fidelity of key patterns during the denoising process, we propose Reflection Sampling (R-Sampling), a step-wise refinement strategy that encourages the reverse trajectory to better align with key prototype constraints. Experiments on nine real-world datasets demonstrate that K-ProtoDiff significantly outperforms state-of-the-art baselines in key pattern retention, achieving an average 77.6% improvement in key pattern preservation.
Generalized Category Discovery (GCD) aims to classify unlabeled data by leveraging knowledge from labeled categories. While existing methods have achieved remarkable progress, they often treat images as flat feature sets, neglecting the intrinsic hierarchy: where key objects dominate meaning and backgrounds serve as context. For instance, in images of a dog either standing on grass or lying on a bed, the dog remains the central semantic element, whereas the background varies. Motivated by this, we propose LEArning Intrinsic Hierarchy (LEAH), a lightweight plug-and-play module designed to model hierarchical structure within images. LEAH consists of two components: a pruner that filters task-irrelevant tokens to extract key objects, and a constructor that embeds key objects and full images into hyperbolic space using adaptive entailment cones to capture compositional semantics. LEAH can be easily integrated into existing GCD frameworks with minimal modification. When applied to SimGCD, it achieves up to 13.2% accuracy improvement on fine-grained benchmarks, demonstrating its effectiveness in discovering subtle inter-class differences through hierarchical modeling.
Incomplete multi-view clustering (IMVC) aims to discover shared cluster structures from multi-view data with partial observations. The core challenges lie in accurately imputing missing views without introducing bias, while maintaining semantic consistency across views and compactness within clusters. To address these challenges, we propose DIMVC-HIA, a novel deep IMVC framework that integrates hierarchical imputation and alignment with four key components: (1) view-specific autoencoders for latent feature extraction, coupled with a view-shared clustering predictor to produce soft cluster assignments; (2) a hierarchical imputation module that first estimates missing cluster assignments based on cross-view contrastive similarity, and then reconstructs missing features using intra-view, intra-cluster statistics; (3) an energy-based semantic alignment module, which promotes intra-cluster compactness by minimizing energy variance around low-energy cluster anchors; and (4) a contrastive assignment alignment module, which enhances cross-view consistency and encourages confident, well-separated cluster predictions. Experiments on benchmarks demonstrate that our framework achieves superior performance under varying levels of missingness.
Concept bottleneck models (CBMs) improve neural network interpretability by introducing an intermediate layer that maps human-understandable concepts to predictions. Recent work has explored the use of vision-language models (VLMs) to automate concept selection and annotation. However, existing VLM-based CBMs typically require full model retraining when new concepts are involved, which limits their adaptability and flexibility in real-world scenarios, especially considering the rapid evolution of vision-language foundation models. To address these issues, we propose Flexible Concept Bottleneck Model (FCBM), which supports dynamic concept adaptation, including complete replacement of the original concept set. Specifically, we design a hypernetwork that generates prediction weights based on concept embeddings, allowing seamless integration of new concepts without retraining the entire model. In addition, we introduce a modified sparsemax module with a learnable temperature parameter that dynamically selects the most relevant concepts, enabling the model to focus on the most informative features. Extensive experiments on five public benchmarks demonstrate that our method achieves accuracy comparable to state-of-the-art baselines with a similar number of effective concepts. Moreover, the model generalizes well to unseen concepts with just a single epoch of fine-tuning, demonstrating its strong adaptability and flexibility.
Machine unlearning (MU) has emerged as a critical tool for removing sensitive or personal information from machine learning models, empowering individuals with the right to be forgotten. While MU has achieved success in classification and generative tasks, whether this technique can be effectively applied to segmentation foundation models remains uncertain. To address this issue, we propose an efficient method, Selective Concept Unlearning (SCU), to unlearn the segmentation capability of target concepts. SCU consists of several key aspects: (1) The Multi-level Forgetting Module, designed with a hierarchical three-level suppression strategy, including (i) distillation-level: Negative distillation steers model’s output distribution away from teacher’s correct outputs, erasing its learned concept recognition. (ii) attention-level: Attention suppression minimizes model’s attention to target regions. (iii) output-level: Directly erases predictions for the target by relabeling as background. (2) The Preservation Module ensures maintaining segmentation quality for non-target concepts. Additionally, we introduce a set of metrics to evaluate segmentation unlearning methods. Experiments demonstrate that SCU consistently outperforms existing baselines.
Graph Neural Networks (GNNs) have emerged as a dominant paradigm for graph classification. Specifically, most existing GNNs mainly rely on the message passing strategy between neighbor nodes, where the expressivity is limited by the 1-dimensional Weisfeiler-Lehman (1-WL) test. Although a number of k-WL-based GNNs have been proposed to overcome this limitation, their computational cost increases rapidly with k, significantly restricting the practical applicability. Moreover, since the k-WL models mainly operate on node tuples, these k-WL-based GNNs cannot retain fine-grained node- or edge-level semantics required by attribution methods (e.g., Integrated Gradients), leading to the less interpretable problem. To overcome the above shortcomings, in this paper, we propose a novel Line Graph Aggregation Network (LGAN), that constructs a line graph from the induced subgraph centered at each node to perform the higher-order aggregation. We theoretically prove that the LGAN not only possesses the greater expressive power than the 2-WL under injective aggregation assumptions, but also has lower time complexity. Empirical evaluations on benchmarks demonstrate that the LGAN outperforms state-of-the-art k-WL-based GNNs, while offering better interpretability.
Graph neural networks (GNNs) have demonstrated impressive performance in a broad spectrum of fields, but always suffer from the generalization problem when confronted with out-of-distribution (OOD) scenarios. Information bottleneck (IB) principle, which endeavors to learn the minimally sufficient representations for downstream tasks, has been shown to be a promising strategy in dealing with this problem. However, the IB-based methods do not inherently distinguish between causal and non-causal parts in the graph, leading to underperforming OOD generalization ability. In this paper, we develop the Graph Causal Information Bottleneck (GCIB) framework, a causal extension of the IB for graph data, which is capable of jointly compressing abundant information and capturing causal dependency from the input graph. Specifically, we endow graph IB with the ability of maintaining causal control by incorporating the underlying causal structure and introducing intervention operation. On this basis, we formulate the learning objective for GCIB and present its specific implementation. Graph representations learned by GCIB can effectively preserve causal information that fundamentally determines graph properties, resulting in outstanding OOD generalization ability. Extensive experiments on both synthetic and real-world datasets demonstrate the superiority of GCIB over state-of-the-art baselines.
Deep learning models such as MLP, Transformer, and TCN have achieved remarkable success in univariate time series forecasting, typically relying on sliding window samples from historical data for training. However, while these models implicitly compress historical information into their parameters during training, they are unable to explicitly and dynamically access this global knowledge during inference, relying only on the local context within the lookback window. This results in an underutilization of rich patterns from the global history. To bridge this gap, we propose Predicting the Future by Retrieving the Past (PFRP), a novel approach that explicitly integrates global historical data to enhance forecasting accuracy. Specifically, we construct a Global Memory Bank (GMB) to effectively store and manage global historical patterns. A retrieval mechanism is then employed to extract similar patterns from the GMB, enabling the generation of global predictions. By adaptively combining these global predictions with the outputs of any local prediction model, PFRP produces more accurate and interpretable forecasts. Extensive experiments conducted on seven real-world datasets demonstrate that PFRP enhances the average performance of advanced univariate forecasting models by 8.4%.
Closer to Biological Mechanism: Drug-Drug Interaction Prediction from the Perspective of Pharmacophore
PDF ↗Drug combinations are widely used in modern medicine but may cause severe adverse drug reactions. Therefore, making effective drug-drug interactions (DDI) prediction is crucial for pharmacovigilance. Existing DDI prediction models are typically built from a structural perspective, assuming that drugs with similar molecular structures may exhibit similar interactions. However, such approaches overlook the biological mechanisms underlying DDI in the human body. This not only weakens the generalization ability of the model, but also makes its interpretability less convincing. Inspired by this, we propose a new method called PC-DDI. Unlike structure-based models, PC-DDI utilizes pharmacophores as basic unit, and designs a complete pharmacophore feature processing framework. It further constructs a pharmacophore-based bipartite graph to model interactions between pharmacophores. This approach allows us to explore the underlying mechanisms of DDI from a functional perspective. We also design a spatial attention weight graph convolution module to optimize the message passing process by integrating pharmacophore position features with node features. Furthermore, we apply causal inference to identify key pharmacophores in pharmacophore bipartite graph, enhancing the interpretability. Compared with the SOTA, PC-DDI achieves an accuracy improvement of 1.84% under the transductive setting and consistently outperforms others in all other experiments.
Estimating counterfactual outcomes from observational data is critical for informed decision-making in domains such as personalized marketing, healthcare, and online platforms. In these contexts, decision processes frequently involve high-dimensional combinatorial interventions, including bundled channel allocation or product set recommendations. For such scenarios, both causal assessment of historical strategies and optimization of novel interventions necessitate models capable of extrapolating to intervention combinations that are underrepresented or entirely absent in observational data. Specifically, in digital marketing, companies often need to evaluate new combinations of channels or target emerging user segments that have not been previously exposed. This challenge is exacerbated by inherent biases in observational datasets, stemming from prior allocation policies and targeting mechanisms, which further aggravate coverage sparsity and compromise off-support counterfactual inference. In this work, we propose Dual-Source Counterfactual Fusion (DSCF), a scalable framework that enables accurate counterfactual prediction under high-dimensional combinatorial interventions, with improved robustness to confounding bias. DSCF jointly models observational data and proxy counterfactual samples through a dual-head mixture-of-experts architecture and domain-guided fusion. This design effectively integrates bias reduction and information diversity while enabling adaptive generalization to counterfactual inputs. Extensive experiments on both synthetic and semi-synthetic datasets demonstrate the effectiveness and robustness of DSCF across diverse scenarios.
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.
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.