[CLS] Token Tells Everything Needed for Training-free Efficient MLLMs

Multimodal Large Language Models (MLLMs) have recently demonstrated strong performance across a wide range of vision-language tasks, garnering significant attention in the computer vision. However, their efficient deployment remains a substantial challenge due to high computational costs and memory requirements. Recognizing the redundancy of information within the vision modality, recent studies have explored methods for compressing visual tokens in MLLMs to enhance efficiency in a training-free manner. Despite their effectiveness, existing methods like Fast rely on the attention between visual tokens and prompt text tokens as the importance indicator, overlooking the relevance to response text and thus introducing perception bias. In this paper, we demonstrate that in MLLMs, the [CLS] token in the visual encoder inherently knows which visual tokens are important for MLLMs. Building on this prior, we introduce a simple yet effective method for train-free visual token compression, called VTC-CLS. Firstly, it leverages the attention score of the [CLS] token on visual tokens as an importance indicator for pruning visual tokens. Besides, we also explore ensembling the importance scores derived by the [CLS] token from different layers to capture the key visual information more comprehensively. Extensive experiments demonstrate that our VTC-CLS achieves the state-of-the-art performance across various tasks compared with baseline methods. It also brings notably less computational costs in a training-free manner, highlighting its effectiveness and superiority. Code and models are available at \url{https://github.com/THU-MIG/VTC-CLS}.

PrefixKV: Adaptive Prefix KV Cache is What Vision Instruction-Following Models Need for Efficient Generation

Recently, large vision-language models (LVLMs) have rapidly gained popularity for their strong generation and reasoning capabilities given diverse multimodal inputs. However, these models incur significant computational and memory overhead during inference, which greatly hinders the efficient deployment in practical scenarios. The extensive key-value (KV) cache, necessitated by the lengthy input and output sequences, notably contributes to the high inference cost. Based on this, recent works have investigated ways to reduce the KV cache size for higher efficiency. Although effective, they generally overlook the distinct importance distributions of KV vectors across layers and maintain the same cache size for each layer during the next token prediction. This results in the significant contextual information loss for certain layers, leading to notable performance decline. To address this, we present PrefixKV. It reframes the challenge of determining KV cache sizes for all layers into the task of searching for the optimal global prefix configuration. With an adaptive layer-wise KV retention recipe based on binary search, the maximum contextual information can thus be preserved in each layer, facilitating the generation. Extensive experiments demonstrate that our method achieves the state-of-the-art performance compared with others. It exhibits superior inference efficiency and generation quality trade-offs, showing promising potential for practical applications. Code is available at \url{https://github.com/THU-MIG/PrefixKV}.

Hybrid Discriminative Attribute-Object Embedding Network for Compositional Zero-Shot Learning

Compositional Zero-Shot Learning (CZSL) recognizes new combinations by learning from known attribute-object pairs. However, the main challenge of this task lies in the complex interactions between attributes and object visual representations, which lead to significant differences in images. In addition, the long-tail label distribution in the real world makes the recognition task more complicated. To address these problems, we propose a novel method, named Hybrid Discriminative Attribute-Object Embedding (HDA-OE) network. To increase the variability of training data, HDA-OE introduces an attribute-driven data synthesis (ADDS) module. ADDS generates new samples with diverse attribute labels by combining multiple attributes of the same object. By expanding the attribute space in the dataset, the model is encouraged to learn and distinguish subtle differences between attributes. To further improve the discriminative ability of the model, HDA-OE introduces the subclass-driven discriminative embedding (SDDE) module, which enhances the subclass discriminative ability of the encoding by embedding subclass information in a fine-grained manner, helping to capture the complex dependencies between attributes and object visual features. The proposed model has been evaluated on three benchmark datasets, and the results verify its effectiveness and reliability.

Relation-Aware Meta-Learning for Zero-shot Sketch-Based Image Retrieval

Sketch-based image retrieval (SBIR) relies on free-hand sketches to retrieve natural photos within the same class. However, its practical application is limited by its inability to retrieve classes absent from the training set. To address this limitation, the task has evolved into Zero-Shot Sketch-Based Image Retrieval (ZS-SBIR), where model performance is evaluated on unseen categories. Traditional SBIR primarily focuses on narrowing the domain gap between photo and sketch modalities. However, in the zero-shot setting, the model not only needs to address this cross-modal discrepancy but also requires a strong generalization capability to transfer knowledge to unseen categories. To this end, we propose a novel framework for ZS-SBIR that employs a pair-based relation-aware quadruplet loss to bridge feature gaps. By incorporating two negative samples from different modalities, the approach prevents positive features from becoming disproportionately distant from one modality while remaining close to another, thus enhancing inter-class separability. We also propose a Relation-Aware Meta-Learning Network (RAMLN) to obtain the margin, a hyper-parameter of cross-modal quadruplet loss, to improve the generalization ability of the model. RAMLN leverages external memory to store feature information, which it utilizes to assign optimal margin values. Experimental results obtained on the extended Sketchy and TU-Berlin datasets show a sharp improvement over existing state-of-the-art methods in ZS-SBIR.

Promptable Anomaly Segmentation with SAM Through Self-Perception Tuning

Segment Anything Model (SAM) has made great progress in anomaly segmentation tasks due to its impressive generalization ability. However, existing methods that directly apply SAM through prompting often overlook the domain shift issue, where SAM performs well on natural images but struggles in industrial scenarios. Parameter-Efficient Fine-Tuning (PEFT) offers a promising solution, but it may yield suboptimal performance by not adequately addressing the perception challenges during adaptation to anomaly images. In this paper, we propose a novel Self-Perceptinon Tuning (SPT) method, aiming to enhance SAM's perception capability for anomaly segmentation. The SPT method incorporates a self-drafting tuning strategy, which generates an initial coarse draft of the anomaly mask, followed by a refinement process. Additionally, a visual-relation-aware adapter is introduced to improve the perception of discriminative relational information for mask generation. Extensive experimental results on several benchmark datasets demonstrate that our SPT method can significantly outperform baseline methods, validating its effectiveness. Models and codes will be available online.

Multi-Stage Knowledge Integration of Vision-Language Models for Continual Learning

Vision Language Models (VLMs), pre-trained on large-scale image-text datasets, enable zero-shot predictions for unseen data but may underperform on specific unseen tasks. Continual learning (CL) can help VLMs effectively adapt to new data distributions without joint training, but faces challenges of catastrophic forgetting and generalization forgetting. Although significant progress has been achieved by distillation-based methods, they exhibit two severe limitations. One is the popularly adopted single-teacher paradigm fails to impart comprehensive knowledge, The other is the existing methods inadequately leverage the multimodal information in the original training dataset, instead they rely on additional data for distillation, which increases computational and storage overhead. To mitigate both limitations, by drawing on Knowledge Integration Theory (KIT), we propose a Multi-Stage Knowledge Integration network (MulKI) to emulate the human learning process in distillation methods. MulKI achieves this through four stages, including Eliciting Ideas, Adding New Ideas, Distinguishing Ideas, and Making Connections. During the four stages, we first leverage prototypes to align across modalities, eliciting cross-modal knowledge, then adding new knowledge by constructing fine-grained intra- and inter-modality relationships with prototypes. After that, knowledge from two teacher models is adaptively distinguished and re-weighted. Finally, we connect between models from intra- and inter-task, integrating preceding and new knowledge. Our method demonstrates significant improvements in maintaining zero-shot capabilities while supporting continual learning across diverse downstream tasks, showcasing its potential in adapting VLMs to evolving data distributions.

A Fresh Look at Generalized Category Discovery through Non-negative Matrix Factorization

Generalized Category Discovery (GCD) aims to classify both base and novel images using labeled base data. However, current approaches inadequately address the intrinsic optimization of the co-occurrence matrix $\bar{A}$ based on cosine similarity, failing to achieve zero base-novel regions and adequate sparsity in base and novel domains. To address these deficiencies, we propose a Non-Negative Generalized Category Discovery (NN-GCD) framework. It employs Symmetric Non-negative Matrix Factorization (SNMF) as a mathematical medium to prove the equivalence of optimal K-means with optimal SNMF, and the equivalence of SNMF solver with non-negative contrastive learning (NCL) optimization. Utilizing these theoretical equivalences, it reframes the optimization of $\bar{A}$ and K-means clustering as an NCL optimization problem. Moreover, to satisfy the non-negative constraints and make a GCD model converge to a near-optimal region, we propose a GELU activation function and an NMF NCE loss. To transition $\bar{A}$ from a suboptimal state to the desired $\bar{A}^*$, we introduce a hybrid sparse regularization approach to impose sparsity constraints. Experimental results show NN-GCD outperforms state-of-the-art methods on GCD benchmarks, achieving an average accuracy of 66.1\% on the Semantic Shift Benchmark, surpassing prior counterparts by 4.7\%.

Part-Whole Relational Fusion Towards Multi-Modal Scene Understanding

Multi-modal fusion has played a vital role in multi-modal scene understanding. Most existing methods focus on cross-modal fusion involving two modalities, often overlooking more complex multi-modal fusion, which is essential for real-world applications like autonomous driving, where visible, depth, event, LiDAR, etc., are used. Besides, few attempts for multi-modal fusion, \emph{e.g.}, simple concatenation, cross-modal attention, and token selection, cannot well dig into the intrinsic shared and specific details of multiple modalities. To tackle the challenge, in this paper, we propose a Part-Whole Relational Fusion (PWRF) framework. For the first time, this framework treats multi-modal fusion as part-whole relational fusion. It routes multiple individual part-level modalities to a fused whole-level modality using the part-whole relational routing ability of Capsule Networks (CapsNets). Through this part-whole routing, our PWRF generates modal-shared and modal-specific semantics from the whole-level modal capsules and the routing coefficients, respectively. On top of that, modal-shared and modal-specific details can be employed to solve the issue of multi-modal scene understanding, including synthetic multi-modal segmentation and visible-depth-thermal salient object detection in this paper. Experiments on several datasets demonstrate the superiority of the proposed PWRF framework for multi-modal scene understanding. The source code has been released on https://github.com/liuyi1989/PWRF.

Context Enhancement with Reconstruction as Sequence for Unified Unsupervised Anomaly Detection

Unsupervised anomaly detection (AD) aims to train robust detection models using only normal samples, while can generalize well to unseen anomalies. Recent research focuses on a unified unsupervised AD setting in which only one model is trained for all classes, i.e., n-class-one-model paradigm. Feature-reconstruction-based methods achieve state-of-the-art performance in this scenario. However, existing methods often suffer from a lack of sufficient contextual awareness, thereby compromising the quality of the reconstruction. To address this issue, we introduce a novel Reconstruction as Sequence (RAS) method, which enhances the contextual correspondence during feature reconstruction from a sequence modeling perspective. In particular, based on the transformer technique, we integrate a specialized RASFormer block into RAS. This block enables the capture of spatial relationships among different image regions and enhances sequential dependencies throughout the reconstruction process. By incorporating the RASFormer block, our RAS method achieves superior contextual awareness capabilities, leading to remarkable performance. Experimental results show that our RAS significantly outperforms competing methods, well demonstrating the effectiveness and superiority of our method. Our code is available at https://github.com/Nothingtolose9979/RAS.

LLMI3D: Empowering LLM with 3D Perception from a Single 2D Image

Recent advancements in autonomous driving, augmented reality, robotics, and embodied intelligence have necessitated 3D perception algorithms. However, current 3D perception methods, particularly small models, struggle with processing logical reasoning, question-answering, and handling open scenario categories. On the other hand, generative multimodal large language models (MLLMs) excel in general capacity but underperform in 3D tasks, due to weak spatial and local object perception, poor text-based geometric numerical output, and inability to handle camera focal variations. To address these challenges, we propose the following solutions: Spatial-Enhanced Local Feature Mining for better spatial feature extraction, 3D Query Token-Derived Info Decoding for precise geometric regression, and Geometry Projection-Based 3D Reasoning for handling camera focal length variations. We employ parameter-efficient fine-tuning for a pre-trained MLLM and develop LLMI3D, a powerful 3D perception MLLM. Additionally, we have constructed the IG3D dataset, which provides fine-grained descriptions and question-answer annotations. Extensive experiments demonstrate that our LLMI3D achieves state-of-the-art performance, significantly outperforming existing methods.