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.

With the advent of large vision-language models (LVLMs) demonstrating increasingly human-like abilities, a pivotal question emerges: do different LVLMs interpret multimodal sarcasm differently, and can a single model grasp sarcasm from multiple perspectives like humans? To explore this, we introduce an analytical framework using systematically designed prompts on existing multimodal sarcasm datasets. Evaluating 12 state-of-the-art LVLMs over 2409 samples, we examine interpretive variations within and across models, focusing on confidence levels, alignment with dataset labels, and recognition of ambiguous “neutral” cases. We further validate our findings on a diverse 100-sample mini-benchmark, incorporating multiple datasets, expanded prompt variants, and representative commercial LVLMs. Our findings reveal notable discrepancies—across LVLMs and within the same model under varied prompts. While classification-oriented prompts yield higher internal consistency, models diverge markedly when tasked with interpretive reasoning. These results challenge binary labeling paradigms by highlighting sarcasm’s subjectivity. We advocate moving beyond rigid annotation schemes toward multiperspective, uncertainty-aware modeling, offering deeper insights into multimodal sarcasm comprehension.

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.

The debate between self-interpretable models and post-hoc explanations for black-box models is central to Explainable AI (XAI). Self-interpretable models, such as concept-based networks, offer insights by connecting decisions to human-understandable concepts but often struggle with performance and scalability. Conversely, post-hoc methods like Shapley values, while theoretically robust, are computationally expensive and resource-intensive. To bridge the gap between these two lines of research, we propose a novel method that combines their strengths, providing theoretically guaranteed self-interpretability for black-box models without compromising prediction accuracy. Specifically, we introduce a parameter-efficient pipeline, AutoGnothi, which integrates a small side network into the black-box model, allowing it to generate Shapley value explanations without changing the original network parameters. This side-tuning approach significantly reduces memory, training, and inference costs, outperforming traditional parameter-efficient methods, where full fine-tuning serves as the optimal baseline. AutoGnothi enables the black-box model to predict and explain its predictions with minimal overhead. Extensive experiments show that AutoGnothi offers accurate explanations for both vision and language tasks, delivering superior computational efficiency with comparable interpretability.

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.

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.

Parameter-efficient transfer learning (PETL) methods show promise in adapting a pre-trained model to various downstream tasks while training only a few parameters. In the computer vision (CV) domain, numerous PETL algorithms have been proposed, but their direct employment or comparison remains inconvenient. To address this challenge, we construct a Unified Visual PETL Benchmark (V-PETL Bench) for the CV domain by selecting 30 diverse, challenging, and comprehensive datasets from image recognition, video action recognition, and dense prediction tasks. On these datasets, we systematically evaluate 25 dominant PETL algorithms and open-source a modular and extensible codebase for fair evaluation of these algorithms. V-PETL Bench runs on NVIDIA A800 GPUs and requires approximately 310 GPU days. We release all the benchmark, making it more efficient and friendly to researchers. Additionally, V-PETL Bench will be continuously updated for new PETL algorithms and CV tasks.