The quadratic complexity of Multimodal Large Language Models (MLLMs) with respect to context length poses significant computational and memory challenges, hindering their real-world deployment. In the paper, we devise a ''filter-correlate-compress'' framework to accelerate the MLLM by systematically optimizing multimodal context length during prefilling. The framework first implements FiCoCo-V, a training-free method operating within the vision encoder. It employs a redundancy-based token discard mechanism that uses a novel integrated metric to accurately filter out redundant visual tokens. To mitigate information loss, the framework introduces a correlation-based information recycling mechanism that allows preserved tokens to selectively recycle information from correlated discarded tokens with a self-preserving compression, thereby preventing the dilution of their own core content. The framework's FiCoCo-L variant further leverages task-aware textual priors to perform token reduction directly within the LLM decoder. Extensive experiments demonstrate that the FiCoCo series effectively accelerates a range of MLLMs, achieves up to 14.7× FLOPs reduction with 93.6% performance retention. Our methods consistently outperform state-of-the-art training-free approaches, showcasing effectiveness and generalizability across model architectures, sizes, and tasks without requiring retraining.

Large vision-language models (LVLMs) excel at visual understanding but face efficiency challenges due to quadratic complexity when processing long multimodal contexts. While token compression can reduce computational costs, existing approaches are designed for single-view LVLMs and fail to account for the unique multi-view characteristics of high-resolution LVLMs that use dynamic cropping. Current methods treat all tokens uniformly, yet our analysis shows that global thumbnails can naturally guide the compression of local crops by providing holistic context for evaluating informativeness. In this paper, we first analyze the dynamic cropping strategy, revealing both the complementary relationship between thumbnails and crops and the distinct characteristics across different crops. Based on these insights, we propose ''Global Compression Commander'' (GlobalCom2), a novel plug-and-play token compression framework for high-resolution LVLMs. GlobalCom2 uses the thumbnail as a ''commander'' to adaptively guide the compression of local crops, preserving informative details while removing redundancy. Extensive experiments demonstrate that GlobalCom2 maintains over 90% of model performance while compressing 90% of visual tokens, reducing FLOPs to 9.1% and peak memory usage to 60% of the original.

Vision-Language-Action (VLA) models typically bridge the gap between perceptual and action spaces by pre-training a large-scale Vision-Language Model (VLM) on robotic data. While this approach greatly enhances performance, it also incurs significant training costs. In this paper, we investigate how to effectively bridge vision-language (VL) representations to action (A). We introduce VLA-Adapter, a novel paradigm designed to reduce the reliance of VLA models on large-scale VLMs and extensive pre-training. To this end, we first systematically analyze the effectiveness of various VL conditions and present key findings on which conditions are essential for bridging perception and action spaces. Based on these insights, we propose a lightweight Policy module with Bridge Attention, which autonomously injects the optimal condition into the action space. In this way, our method achieves high performance using only a 0.5B-parameter backbone, without any robotic data pre-training. Extensive experiments on both simulated and real-world robotic benchmarks show that VLA-Adapter not only achieves state-of-the-art level performance, but also offers the fast inference speed reported to date. Furthermore, thanks to the proposed advanced bridging paradigm, VLA-Adapter enables the training of a powerful VLA model on a single consumer-grade GPU, greatly lowering the barrier to deploying VLA model.

A dexterous hand capable of generalizable grasping objects is fundamental for the development of general-purpose embodied AI. However, previous methods focus narrowly on low-level grasp stability metrics, neglecting affordance-aware positioning and human-like poses which are crucial for downstream manipulation. To address these limitations, we propose AffordDex, a novel framework with two-stage training that learns a universal grasping policy with an inherent understanding of both motion priors and object affordances. In the first stage, a trajectory imitator is pre-trained on a large corpus of human hand motions to instill a strong prior for natural movement. In the second stage, a residual module is trained to adapt these general human-like motions to specific object instances. This refinement is critically guided by two components: our Negative Affordance-aware Segmentation (NAA) module, which identifies functionally inappropriate contact regions, and a privileged teacher-student distillation process that ensures the final vision-based policy is highly successful. Extensive experiments demonstrate that AffordDex not only achieves universal dexterous grasping but also remains remarkably human-like in posture and functionally appropriate in contact location. As a result, AffordDex significantly outperforms state-of-the-art baselines across seen objects, unseen instances, and even entirely novel categories.

Referring expression comprehension (REC) is a vision-language task to locate a target object in an image based on a language expression. Fully fine-tuning general-purpose pre-trained vision-language foundation models for REC yields impressive performance but becomes increasingly costly. Parameter-efficient transfer learning (PETL) methods have shown strong performance with fewer tunable parameters. However, directly applying PETL to REC faces two challenges: (1) insufficient multi-modal interaction between pre-trained vision-language foundation models, and (2) high GPU memory usage due to gradients passing through the heavy vision-language foundation models. To this end, we present M2IST: Multi-Modal Interactive Side-Tuning with M3ISAs: Mixture of Multi-Modal Interactive Side-Adapters. During fine-tuning, we fix the pre-trained uni-modal encoders and update M3ISAs to enable efficient vision-language alignment for REC. Empirical results reveal that M2IST achieves better performance-efficiency trade-off than full fine-tuning and other PETL methods, requiring only 2.11% tunable parameters, 39.61% GPU memory, and 63.46% training time while maintaining competitive performance. Our code is released at https://github.com/xuyang-liu16/M2IST.

In recent years, applying multi-modal large language models (MLLMs) in various fields has achieved remarkable success. However, as the foundation model for many downstream tasks, MLLMs comprise the well-known Transformer network, which has a less efficient quadratic computation complexity. In this study, we introduce Cobra, a multi-modal large-scale language model built upon a state-space model, which has demonstrated significant potential in efficiently handling long sequences with fast inference and linear scalability concerning sequence length. Specifically, Cobra involves replacing Transformer-based backbone models (e.g., LLaMA or Phi) with pre-trained Mamba language models. We then empirically explore effective strategies for aligning visual and textual modalities and integrating various pre-trained Mamba model variants with visual encoders. Experiments across various multi-modal benchmarks demonstrate that: (i) Cobra performs 3× ∼ 4× faster than the most computationally efficient state-of-the-art methods, e.g., LLaVA-Phi and MobileVLM v2. Additionally, its performance is significantly enhanced thanks to the implementation of linear sequential modeling. (ii) Cobra fine-tunes a small parameter (∼48% of model parameters), leading to a significant improvement in overall performance compared to LLaVA.

In robotic visuomotor policy learning, diffusion-based models have achieved significant success in improving the accuracy of action trajectory generation compared to traditional autoregressive models. However, they suffer from inefficiency due to multiple denoising steps and limited flexibility from complex constraints. In this paper, we introduce Coarse-to-Fine AutoRegressive Policy (CARP), a novel paradigm for visuomotor policy learning that redefines the autoregressive action generation process as a coarse-to-fine, next-scale approach. CARP decouples action generation into two stages: first, an action autoencoder learns multi-scale representations of the entire action sequence; then, a GPT-style transformer refines the sequence prediction through a coarse-to-fine autoregressive process. This straightforward and intuitive approach produces highly accurate and smooth actions, matching or even surpassing the performance of diffusion-based policies while maintaining efficiency on par with autoregressive policies. We conduct extensive evaluations across diverse settings, including single-task and multi-task scenarios on state-based and image-based simulation benchmarks, as well as real-world tasks. CARP achieves competitive success rates, with up to a 10% improvement, and delivers 10x faster inference compared to state-of-the-art policies, establishing a high-performance, efficient, and flexible paradigm for action generation in robotic tasks.

Diffusion transformers have shown significant effectiveness in both image and video synthesis at the expense of huge computation costs. To address this problem, feature caching methods have been introduced to accelerate diffusion transformers by caching the features in previous timesteps and reusing them in the following timesteps. However, previous caching methods ignore that different tokens exhibit different sensitivities to feature caching, and feature caching on some tokens may lead to 10$\times$ more destruction to the overall generation quality compared with other tokens. In this paper, we introduce token-wise feature caching, allowing us to adaptively select the most suitable tokens for caching, and further enable us to apply different caching ratios to neural layers in different types and depths. Extensive experiments on PixArt-alpha, OpenSora, and DiT demonstrate our effectiveness in both image and video generation with no requirements for training. For instance, 2.36$\times$ and 1.93$\times$ acceleration are achieved on OpenSora and PixArt-$\alpha$ with almost no drop in generation quality. Codes have been released in the supplementary material and Github.

This paper addresses the inherent inference latency challenges associated with deploying multimodal large language models (MLLM) in quadruped vision-language-action (QUAR-VLA) tasks. Our investigation reveals that conventional parameter reduction techniques ultimately impair the performance of the language foundation model during the action instruction tuning phase, making them unsuitable for this purpose. We introduce a novel latency-free quadruped MLLM model, dubbed QUARTOnline, designed to enhance inference efficiency without degrading the performance of the language foundation model. By incorporating Action Chunk Discretization (ACD), we compress the original action representation space, mapping continuous action values onto a smaller set of discrete representative vectors while preserving critical information. Subsequently, we fine-tune the MLLM to integrate vision, language, and compressed actions into a unified semantic space. Experimental results demonstrate that QUART-Online operates in tandem with the existing MLLM system, achieving real-time inference at 50 Hz in sync with the underlying controller frequency, significantly boosting the success rate across various tasks by 65 %. Our project page is https://quart-online.github.io.

Despite impressive advancements in Visual-Language Models (VLMs) for multi-modal tasks, their reliance on RGB inputs limits precise spatial understanding. Existing methods for integrating spatial cues, such as point clouds or depth, either require specialized sensors or fail to effectively exploit depth information for higher-order reasoning. To this end, we propose a novel Spatial Sense and Reasoning method, dubbed SSR, a novel framework that transforms raw depth data into structured, interpretable textual rationales. These textual rationales serve as meaningful intermediate representations to significantly enhance spatial reasoning capabilities. Additionally, we leverage knowledge distillation to compress the generated rationales into compact latent embeddings, which facilitate resource-efficient and plug-and-play integration into existing VLMs without retraining. To enable comprehensive evaluation, we introduce a new dataset named SSR-CoT, a million-scale visual-language reasoning dataset enriched with intermediate spatial reasoning annotations, and present SSRBench, a comprehensive multi-task benchmark. Extensive experiments on multiple benchmarks demonstrate SSR substantially improves depth utilization and enhances spatial reasoning, thereby advancing VLMs toward more human-like multi-modal understanding. Project page: [https://yliu-cs.github.io/SSR](https://yliu-cs.github.io/SSR).

Recently, despite the unprecedented success of large pre-trained visual-language models (VLMs) on a wide range of downstream tasks, the real-world unsupervised domain adaptation (UDA) problem is still not well explored. Therefore, in this paper, we first experimentally demonstrate that the unsupervised-trained VLMs can significantly reduce the distribution discrepancy between source and target domains, thereby improving the performance of UDA. However, a major challenge for directly deploying such models on downstream UDA tasks is prompt engineering, which requires aligning the domain knowledge of source and target domains, since the performance of UDA is severely influenced by a good domain-invariant representation. We further propose a Prompt-based Distribution Alignment (PDA) method to incorporate the domain knowledge into prompt learning. Specifically, PDA employs a two-branch prompt-tuning paradigm, namely base branch and alignment branch. The base branch focuses on integrating class-related representation into prompts, ensuring discrimination among different classes. To further minimize domain discrepancy, for the alignment branch, we construct feature banks for both the source and target domains and propose image-guided feature tuning (IFT) to make the input attend to feature banks, which effectively integrates self-enhanced and cross-domain features into the model. In this way, these two branches can be mutually promoted to enhance the adaptation of VLMs for UDA. We conduct extensive experiments on three benchmarks to demonstrate that our proposed PDA achieves state-of-the-art performance. The code is available at https://github.com/BaiShuanghao/Prompt-based-Distribution-Alignment.

Diffusion models have recently achieved remarkable progress in generating realistic images. However, chal-lenges remain in accurately understanding and synthesizing the layout requirements in the textual prompts. To align the generated image with layout instructions, we present a training-free layout calibration system S imM that inter-venes in the generative process on the fly during infer-ence time. Specifically, following a “check-locate-rectify” pipeline, the system first analyses the prompt to generate the target layout and compares it with the intermediate outputs to automatically detect errors. Then, by moving the located activations and making intra- and inter-map adjustments, the rectification process can be performed with negligible computational overhead. To evaluate SimM over a range of layout requirements, we present a benchmark SimMBench that compensates for the lack of superlative spatial relations in existing datasets. And both quantitative and qualitative results demonstrate the effectiveness of the proposed SimM in calibrating the layout inconsistencies. Our project page is at https://simm-t2i.github.io/SimM.

This study focuses on a novel task in text-to-image (T2I) generation, namely action customization. The objective of this task is to learn the co-existing action from limited data and generalize it to unseen humans or even animals. Experimental results show that existing subject-driven customization methods fail to learn the representative characteristics of actions and struggle in decoupling actions from context features, including appearance. To overcome the preference for low-level features and the entanglement of high-level features, we propose an inversion-based method Action-Disentangled Identifier (ADI) to learn action-specific identifiers from the exemplar images. ADI first expands the semantic conditioning space by introducing layer-wise identifier tokens, thereby increasing the representational richness while distributing the inversion across different features. Then, to block the inversion of action-agnostic features, ADI extracts the gradient invariance from the constructed sample triples and masks the updates of irrelevant channels. To comprehensively evaluate the task, we present an Action-Bench that includes a variety of actions, each accompanied by meticulously selected samples. Both quantitative and qualitative results show that our ADI outperforms existing baselines in action-customized T2I generation. Our project page is at https://adi-t2i.github.io/ADI.

Recent compositional zero-shot learning (CZSL) methods adapt pre-trained vision-language models (VLMs) by constructing trainable prompts only for composed state-object pairs. Relying on learning the joint representation of seen compositions, these methods ignore the explicit modeling of the state and object, thus limiting the exploitation of pre-trained knowledge and generalization to unseen compositions. With a particular focus on the universality of the solution, in this work, we propose a novel paradigm for CZSL models that establishes three identification branches (i.e., Multi-Path) to jointly model the state, object, and composition. The presented Troika is an outstanding implementation that aligns the branch-specific prompt representations with decomposed visual features. To calibrate the bias between semantically similar multi-modal representations, we further devise a Cross-Modal Traction module into Troika that shifts the prompt representation towards the current visual content. We conduct extensive experiments on three popular benchmarks, where our method significantly outperforms existing methods in both closed-world and open-world settings. The code will be available at https://github.com/bighuang624/Troika.

Large-scale text-to-image diffusion models have shown impressive capabilities for generative tasks by leveraging strong vision-language alignment from pre-training. However, most vision-language discriminative tasks require extensive fine-tuning on carefully-labeled datasets to acquire such alignment, with great cost in time and computing resources. In this work, we explore directly applying a pre-trained generative diffusion model to the challenging discriminative task of visual grounding without any fine-tuning and additional training dataset. Specifically, we propose VGDiffZero, a simple yet effective zero-shot visual grounding framework based on text-to-image diffusion models. We also design a comprehensive region-scoring method considering both global and local contexts of each isolated proposal. Extensive experiments on RefCOCO, RefCOCO+, and RefCOCOg show that VGDiffZero achieves strong performance on zero-shot visual grounding. Our code is available at https://github.com/xuyang-liu16/VGDiffZero.

Visual grounding (VG) is a challenging task to localize an object in an image based on a textual description. Recent surge in the scale of VG models has substantially improved performance, but also introduced a significant burden on computational costs during fine-tuning. In this paper, we explore applying parameter-efficient transfer learning (PETL) to efficiently transfer the pre-trained vision-language knowledge to VG. Specifically, we propose DARA, a novel PETL method comprising Domain-aware Adapters (DA Adapters) and Relation-aware Adapters (RA Adapters) for VG. DA Adapters first transfer intra-modality representations to be more fine-grained for the VG domain. Then RA Adapters share weights to bridge the relation between two modalities, improving spatial reasoning. Empirical results on widely-used benchmarks demonstrate that DARA achieves the best accuracy while saving numerous updated parameters compared to the full fine-tuning and other PETL methods. Notably, with only 2.13% tunable backbone parameters, DARA improves average accuracy by 0.81% across the three benchmarks compared to the baseline model. Our code is available at https://github.com/liuting20/DARA.

To address the occlusion issues in person Re-Identification (ReID) tasks, many methods have been proposed to extract part features by introducing external spatial information. However, due to missing part appearance information caused by occlusion and noisy spatial information from external model, these purely vision-based approaches fail to correctly learn the features of human body parts from limited training data and struggle in accurately locating body parts, ultimately leading to misaligned part features. To tackle these challenges, we propose a Prompt-guided Feature Disentangling method (ProFD), which leverages the rich pre-trained knowledge in the textual modality facilitate model to generate well-aligned part features. ProFD first designs part-specific prompts and utilizes noisy segmentation mask to preliminarily align visual and textual embedding, enabling the textual prompts to have spatial awareness. Furthermore, to alleviate the noise from external masks, ProFD adopts a hybrid-attention decoder, ensuring spatial and semantic consistency during the decoding process to minimize noise impact. Additionally, to avoid catastrophic forgetting, we employ a self-distillation strategy, retaining pre-trained knowledge of CLIP to mitigate over-fitting. Evaluation results on the Market1501, DukeMTMC-ReID, Occluded-Duke, Occluded-ReID, and P-DukeMTMC datasets demonstrate that ProFD achieves state-of-the-art results.

Many recent studies leverage the pre-trained CLIP for text-video cross-modal retrieval by tuning the backbone with additional heavy modules, which not only brings huge computational burdens with much more parameters, but also leads to the knowledge forgetting from upstream models. In this work, we propose the VoP: Text-Video Cooperative Prompt Tuning for efficient tuning on the text-video retrieval task. The proposed VoP is an end-to-end framework with both video & text prompts introducing, which can be regarded as a powerful baseline with only 0.1% trainable parameters. Further, based on the spatiotemporal characteristics of videos, we develop three novel video prompt mechanisms to improve the performance with different scales of trainable parameters. The basic idea of the VoP enhancement is to model the frame position, frame context, and layer function with specific trainable prompts, respectively. Extensive experiments show that compared to full fine-tuning, the enhanced VoP achieves a 1.4% average R@1 gain across five text-video retrieval benchmarks with 6× less parameter overhead. The code will be available at https://github.com/bighuang624/VoP.

Compositional zero-shot learning (CZSL) refers to recognizing unseen compositions of known visual primitives, which is an essential ability for artificial intelligence systems to learn and understand the world. While considerable progress has been made on existing benchmarks, we suspect whether popular CZSL methods can address the challenges of few-shot and few referential compositions, which is common when learning in real-world unseen environments. To this end, we study the challenging reference-limited compositional zero-shot learning (RL-CZSL) problem in this paper, i.e., given limited seen compositions that contain only a few samples as reference, unseen compositions of observed primitives should be identified. We propose a novel Meta Compositional Graph Learner (MetaCGL) that can efficiently learn the compositionality from insufficient referential information and generalize to unseen compositions. Besides, we build a benchmark with two new large-scale datasets that consist of natural images with diverse compositional labels, providing more realistic environments for RL-CZSL. Extensive experiments in the benchmarks show that our method achieves state-of-the-art performance in recognizing unseen compositions when reference is limited for compositional learning.

In few-shot image classification scenarios, meta-learning methods aim to learn transferable feature representations extracted from seen domains (base classes) in the meta-training phase and quickly adapt to unseen domains (novel classes) in the meta-testing phase. However, when seen and unseen domains have a large discrepancy, existing approaches do not perform well due to the incapability of generalizing to unseen domains. In this paper, we investigate the challenging domain generalized few-shot image classification problem. We design an Meta Regularization Network (MRN) to learn a domain-invariant discriminative feature space, where a learning to learn update strategy is used to simulate domain shifts caused by seen and unseen domains. The simulation trains the model to learn to reorganize the feature knowledge acquired from seen domains to represent unseen domains. Extensive experiments and analysis show that our proposed MRN can significantly improve the generalization ability of various meta-learning methods to achieve state-of-the-art performance in domain generalized few-shot learning.

The purpose of few-shot recognition is to recognize novel categories with a limited number of labeled examples in each class. To encourage learning from a supplementary view, recent approaches have introduced auxiliary semantic modalities into effective metric-learning frameworks that aim to learn a feature similarity between training samples (support set) and test samples (query set). However, these approaches only augment the representations of samples with available semantics while ignoring the query set, which loses the potential for the improvement and may lead to a shift between the modalities combination and the pure-visual representation. In this paper, we devise an attributes-guided attention module (AGAM) to utilize human-annotated attributes and learn more discriminative features. This plug-and-play module enables visual contents and corresponding attributes to collectively focus on important channels and regions for the support set. And the feature selection is also achieved for query set with only visual information while the attributes are not available. Therefore, representations from both sets are improved in a fine-grained manner. Moreover, an attention alignment mechanism is proposed to distill knowledge from the guidance of attributes to the pure-visual branch for samples without attributes. Extensive experiments and analysis show that our proposed module can significantly improve simple metric-based approaches to achieve state-of-the-art performance on different datasets and settings.

While few-shot learning (FSL) aims for rapid generalization to new concepts with little supervision, self-supervised learning (SSL) constructs supervisory signals directly computed from unlabeled data. Exploiting the complementarity of these two manners, few-shot auxiliary learning has recently drawn much attention to deal with few labeled data. Previous works benefit from sharing inductive bias between the main task (FSL) and auxiliary tasks (SSL), where the shared parameters of tasks are optimized by minimizing a linear combination of task losses. However, it is challenging to select a proper weight to balance tasks and reduce task conflict. To handle the problem as a whole, we propose a novel approach named as Pareto self-supervised training (PSST) for FSL. PSST explicitly decomposes the few-shot auxiliary problem into multiple constrained multi-objective subproblems with different trade-off preferences, and here a preference region in which the main task achieves the best performance is identified. Then, an effective preferred Pareto exploration is proposed to find a set of optimal solutions in such a preference region. Extensive experiments on several public benchmark datasets validate the effectiveness of our approach by achieving state-of-the-art performance.

Few-shot learning aims to generalize to novel classes. It has achieved great success in image and text classification tasks. Inspired by such success, few-shot node classification in homogeneous graph has attracted much attention but few works have begun to study this problem in Heterogeneous Information Network (HIN) so far. We consider few-shot learning in HIN and study a pioneering problem HIN Few-Shot Node Classification (HIN-FSNC) that aims to generalize the node types with sufficient labeled samples to unseen node types with only few-labeled samples. However, existing HIN datasets contain just one labeled node type, which means they cannot meet the setting of unseen node types. To facilitate the investigation of HIN-FSNC, we propose a large-scale academic HIN dataset called HINFShot. It contains 1,235,031 nodes with four node types (author, paper, venue, institution) and all the nodes regardless of node type are divided into 80 classes. Finally, we conduct extensive experiments on HINFShot and the result indicates a significant challenge of identifying novel classes of unseen node types in HIN-FSNC.

Multivariate time series forecasting has attracted wide attention in areas, such as system, traffic, and finance. The difficulty of the task lies in that traditional methods fail to capture complicated non-linear dependencies between time steps and between multiple time series. Recently, recurrent neural network and attention mechanism have been used to model periodic temporal patterns across multiple time steps. However, these models fit not well for time series with dynamic-period patterns or nonperiodic patterns. In this paper, we propose a dual self-attention network (DSANet) for highly efficient multivariate time series forecasting, especially for dynamic-period or nonperiodic series. DSANet completely dispenses with recurrence and utilizes two parallel convolutional components, called global temporal convolution and local temporal convolution, to capture complex mixtures of global and local temporal patterns. Moreover, DSANet employs a self-attention module to model dependencies between multiple series. To further improve the robustness, DSANet also integrates a traditional autoregressive linear model in parallel to the non-linear neural network. Experiments on real-world multivariate time series data show that the proposed model is effective and outperforms baselines.