CVPR 2025 Papers

In robotic manipulation, task goals can be conveyed through various modalities, such as language, goal images, and goal videos. However, natural language can be ambiguous, while images or videos may offer overly detailed specifications. To address these challenges, we propose a novel approach using comprehensive multi-modal prompts that explicitly convey both low-level actions and high-level planning in a simple manner. Specifically, for each key-frame in the task sequence, our method allows for manual or automatic generation of simple and expressive 2D visual prompts overlaid on RGB images. These prompts represent the required task goals, such as the end-effector pose and the desired movement direction after contact. We develop a training strategy that enables the model to interpret these visual-language prompts and predict the corresponding contact poses and movement directions in SE(3) space. Furthermore, by sequentially executing all key-frame steps, the model can complete long-horizon tasks. This approach not only helps the model explicitly understand the task objectives but also enhances its robustness on unseen tasks by providing easily interpretable prompts. We evaluate our method in both simulated and real-world environments, demonstrating its robust manipulation capabilities.

Test-time prompt tuning for vision-language models (VLMs) is getting attention because of their ability to learn with unlabeled data without fine-tuning. Although test-time prompt tuning methods for VLMs can boost accuracy, the resulting models tend to demonstrate poor calibration, which casts doubts on the reliability and trustworthiness of these models. Notably, more attention needs to be devoted to calibrating the test-time prompt tuning in vision-language models. To this end, we propose a new approach, called O-TPT that introduces orthogonality constraints on the textual features corresponding to the learnable prompts for calibrating test-time prompt tuning in VLMs. Towards introducing orthogonality constraints, we make the following contributions. First, we uncover new insights behind the suboptimal calibration performance of existing methods relying on textual feature dispersion. Second, we show that imposing a simple orthogonalization of textual features is a more effective approach towards obtaining textual dispersion. We conduct extensive experiments on various datasets with different backbones and baselines. The results indicate that our method consistently outperforms the prior state of the art in significantly reducing the overall average calibration error. Also, our method surpasses the zero-shot calibration performance on fine-grained classification tasks. Our code is available at https://github.com/ashshaksharifdeen/O-TPT.

Recent studies have shown that large vision-language models (LVLMs) often suffer from the issue of object hallucinations (OH). To mitigate this issue, we introduce an efficient method that edits the model weights based on an unsafe subspace, which we call HalluSpace in this paper. With truthful and hallucinated text prompts accompanying the visual content as inputs, the HalluSpace can be identified by extracting the hallucinated embedding features and removing the truthful representations in LVLMs. By orthogonalizing the model weights, input features will be projected into the Null space of the HalluSpace to reduce OH, based on which we name our method Nullu. We reveal that HalluSpaces generally contain prior information in the large language models (LLMs) applied to build LVLMs, which have been shown as essential causes of OH in previous studies. Therefore, null space projection suppresses the LLMs' priors to filter out the hallucinated features, resulting in contextually accurate outputs. Experiments show that our method can effectively mitigate OH across different LVLM families without extra inference costs and also show strong performance in general LVLM benchmarks. Code is released at https://github.com/Ziwei-Zheng/Nullu.

Vision-Language Models (VLMs) learn a shared feature space for text and images, enabling the comparison of inputs of different modalities. While prior works demonstrated that VLMs organize natural language representations into regular structures encoding composite meanings, it remains unclear if compositional patterns also emerge in the visual embedding space. In this work, we investigate compositionality in the image domain, where the analysis of compositional properties is challenged by noise and sparsity of visual data. We address these problems and propose a framework, called Geodesically Decomposable Embeddings (GDE), that approximates image representations with geometry-aware compositional structures in the latent space. We demonstrate that visual embeddings of pre-trained VLMs exhibit a compositional arrangement, and evaluate the effectiveness of this property in the tasks of compositional classification and group robustness. GDE achieves stronger performance in compositional classification compared to its counterpart method that assumes linear geometry of the latent space. Notably, it is particularly effective for group robustness, where we achieve higher results than task-specific solutions. Our results indicate that VLMs can automatically develop a human-like form of compositional reasoning in the visual domain, making their underlying processes more interpretable. Code is available at https://github.com/BerasiDavide/vlm_image_compositionality.

The emergence of vision-language foundation models, such as CLIP, has revolutionized image-text representation, enabling a broad range of applications via prompt learning. Despite its promise, real-world datasets often contain noisy labels that can degrade prompt learning performance. In this paper, we demonstrate that using mean absolute error (MAE) loss in prompt learning, named PromptMAE, significantly enhances robustness against noisy labels while maintaining high accuracy. Though MAE is straightforward and recognized for its robustness, it is rarely used in noisy-label learning due to its slow convergence and poor performance outside prompt learning scenarios. To elucidate the robustness of PromptMAE, we leverage feature learning theory to show that MAE can suppress the influence of noisy samples, thereby improving the signal-to-noise ratio and enhancing overall robustness. Additionally, we introduce PromptOT, a prompt-based optimal transport data purification method to enhance the robustness further. PromptOT employs text encoder representations in vision-language models as prototypes to construct an optimal transportation matrix. This matrix effectively partitions datasets into clean and noisy subsets, allowing for the application of cross-entropy loss to the clean subset and MAE loss to the noisy subset. Our Noise-Label Prompt Learning method, named NLPrompt, offers a simple and efficient approach that leverages the expressive representation and precise alignment capabilities of vision-language models for robust prompt learning. We validate NLPrompt through extensive experiments across various noise settings, demonstrating significant performance improvements.

We introduce a novel method for pre-training of large-scale vision encoders. Building on recent advancements in autoregressive pre-training of vision models, we extend this framework to a multimodal setting, i.e., images and text. In this paper, we present AIMV2, a family of generalist vision encoders characterized by a straightforward pre-training process, scalability, and remarkable performance across a range of downstream tasks. This is achieved by pairing the vision encoder with a multimodal decoder that autoregressively generates raw image patches and text tokens. Our encoders excel not only in multimodal evaluations but also in vision benchmarks such as localization, grounding, and classification. Notably, our AIMV2-3B encoder achieves 89.5% accuracy on ImageNet-1k with a frozen trunk. Fur- thermore, AIMV2 consistently outperforms state-of-the-art contrastive models (e.g., CLIP, SigLIP) in multimodal im- age understanding across diverse settings.

Today's most advanced vision-language models (VLMs) remain proprietary. The strongest open-weight models rely heavily on synthetic data from proprietary VLMs to achieve good performance, effectively distilling these closed VLMs into open ones. As a result, the community has been missing foundational knowledge about how to build performant VLMs from scratch. We present Molmo, a new family of VLMs that are state-of-the-art in their class of openness. Our key contribution is a collection of new datasets called PixMo, including a dataset of highly detailed image captions for pre-training, a free-form image Q&A dataset for fine-tuning, and an innovative 2D pointing dataset, all collected without the use of external VLMs. The success of our approach relies on careful modeling choices, a well-tuned training pipeline, and, most critically, the quality of our newly collected datasets. Our best-in-class 72B model not only outperforms others in the class of open weight and data models, but also outperforms larger proprietary models including Claude 3.5 Sonnet, and Gemini 1.5 Pro and Flash, second only to GPT-4o based on both academic benchmarks and on a large human evaluation. Our model weights, new datasets, and source code are available at https://molmo.allenai.org/blog.

Error detection in procedural activities is essential for consistent and correct outcomes in AR-assisted and robotic systems. Existing methods often focus on temporal ordering errors or rely on static prototypes to represent normal actions. However, these approaches typically overlook the common scenario where multiple, distinct actions are valid following a given sequence of executed actions. This leads to two issues: (1) the model cannot effectively detect errors using static prototypes when the inference environment or action execution distribution differs from training; and (2) the model may also use the wrong prototypes to detect errors if the ongoing action label is not the same as the predicted one. To address this problem, we propose an Adaptive Multiple Normal Action Representation (AMNAR) framework. AMNAR predicts all valid next actions and reconstructs their corresponding normal action representations, which are compared against the ongoing action to detect errors. Extensive experiments demonstrate that AMNAR achieves state-of-the-art performance, highlighting the effectiveness of AMNAR and the importance of modeling multiple valid next actions in error detection. The code is available at https://github.com/iSEE-Laboratory/AMNAR.

Mobile manipulation is the fundamental challenge for robotics to assist humans with diverse tasks and environments in everyday life. However, conventional mobile manipulation approaches often struggle to generalize across different tasks and environments because of the lack of large-scale training.In contrast, recent advances in vision-language-action (VLA) models have shown impressive generalization capabilities, but these foundation models are developed for fixed-base manipulation tasks.Therefore, we propose an efficient policy adaptation framework to transfer pre-trained VLA models of fix-base manipulation to mobile manipulation, so that high generalization ability across tasks and environments can be achieved in mobile manipulation policy.Specifically, we utilize pre-trained VLA models to generate waypoints of the end-effector with high generalization ability. We design motion planning objectives for the mobile base and the robot arm, which aim at maximizing the physical feasibility of the trajectory. Finally, we present an efficient bi-level objective optimization framework for trajectory generation, where the upper-level optimization predicts waypoints for base movement to enhance the manipulator policy space, and the lower-level optimization selects the optimal end-effector trajectory to complete the manipulation task. Extensive experimental results on OVMM and the real world demonstrate that our method achieves a 4.2% higher success rate than the state-of-the-art mobile manipulation, and only requires 50 training cost for real world deployment due to the strong generalization ability in the pre-trained VLA models.

Despite great success across various multimodal tasks, Large Vision-Language Models (LVLMs) often encounter object hallucinations with generated textual responses being inconsistent with the actual objects in images. We examine different LVLMs and pinpoint that one root cause of object hallucinations lies with deficient attention on discriminative image features. Specifically, LVLMs often predominantly attend to prompt-irrelevant global features instead of prompt-relevant local features, undermining their visual grounding capacity and leading to object hallucinations. We propose Assembly of Global and Local Attention (AGLA) , a training-free and plug-and-play approach that mitigates hallucinations by assembling global features for response generation and local features for visual discrimination simultaneously. Specifically, we introduce an image-prompt matching scheme that captures prompt-relevant local features from images, leading to an augmented view of the input image where prompt-relevant content is highlighted while irrelevant distractions are suppressed. Hallucinations can thus be mitigated with a calibrated logit distribution that is from generative global features of the original image and discriminative local features of the augmented image. Extensive experiments show the superiority of AGLA in LVLM hallucination mitigation, demonstrating its wide applicability across both discriminative and generative tasks. Our code is available at https://github.com/Lackel/AGLA.

Hallucination remains a major challenge for Large Vision-Language Models (LVLMs). Direct Preference Optimization (DPO) has gained increasing attention as a simple solution to hallucination issues. It directly learns from constructed preference pairs that reflect the severity of hallucinations in responses to the same prompt and image. Nonetheless, different data construction methods in existing works bring notable performance variations. We identify a crucial factor here: outcomes are largely contingent on whether the constructed data aligns on-policy w.r.t the initial (reference) policy of DPO. Theoretical analysis suggests that learning from off-policy data is impeded by the presence of KL-divergence between the updated policy and the reference policy. From the perspective of dataset distribution, we systematically summarize the inherent flaws in existing algorithms that employ DPO to address hallucination issues. To alleviate the problems, we propose On-Policy Alignment (OPA)-DPO framework, which uniquely leverages expert feedback to correct hallucinated responses and aligns both the original and expert-revised responses in an on-policy manner. Notably, with only 4.8k data, OPA-DPO achieves an additional reduction in the hallucination rate of LLaVA-1.5-7B: 13.26% on the AMBER benchmark and 5.39% on the Object-Hal benchmark, compared to the previous SOTA algorithm trained with 16k samples.

We propose a novel hybrid Mamba-Transformer backbone, MambaVision, specifically tailored for vision applications. Our core contribution includes redesigning the Mamba formulation to enhance its capability for efficient modeling of visual features. Through a comprehensive ablation study, we demonstrate the feasibility of integrating Vision Transformers (ViT) with Mamba. Our results show that equipping the Mamba architecture with self-attention blocks in the final layers greatly improves its capacity to capture long-range spatial dependencies. Based on these findings, we introduce a family of MambaVision models with a hierarchical architecture to meet various design criteria. For classification on the ImageNet-1K dataset, MambaVision variants achieve state-of-the-art (SOTA) performance in terms of both Top-1 accuracy and throughput. In downstream tasks such as object detection, instance segmentation, and semantic segmentation on MS COCO and ADE20K datasets, MambaVision outperforms comparably sized backbones while demonstrating favorable performance. Code: https://github.com/NVlabs/MambaVision

Mamba, an architecture with RNN-like token mixer of state space model (SSM), was recently introduced to address the quadratic complexity of the attention mechanism and subsequently applied to vision tasks. Nevertheless, the performance of Mamba for vision is often underwhelming when compared with convolutional and attention-based models. In this paper, we delve into the essence of Mamba, and conceptually conclude that Mamba is ideally suited for tasks with long-sequence and autoregressive characteristics. For vision tasks, as image classification on ImageNet does not align with either characteristic, we hypothesize that Mamba is not necessary for this task; Detection and segmentation tasks on COCO or ADE20K are also not autoregressive, yet they adhere to the long-sequence characteristic, so we believe it is still worthwhile to explore Mamba's potential for these tasks. To empirically verify our hypotheses, we construct a series of models named MambaOut through stacking Mamba blocks while removing their core token mixer, SSM. Experimental results strongly support our hypotheses. Specifically, our MambaOut model surpasses all visual Mamba models on ImageNet image classification, indicating that Mamba is indeed unnecessary for this task. As for detection and segmentation, MambaOut cannot match the performance of state-of-the-art visual Mamba models, demonstrating the potential of Mamba for long-sequence visual tasks.

Similar to Vision Transformers, this paper identifies artifacts also present within the feature maps of Vision Mamba. These artifacts, corresponding to high-norm tokens emerging in low-information background areas of images, appear much more severe in Vision Mamba---they exist prevalently even with the tiny-sized model and activate extensively across background regions. To mitigate this issue, we follow the prior solution of introducing register tokens into Vision Mamba. To better cope with Mamba blocks' uni-directional inference paradigm, two key modifications are introduced: 1) evenly inserting registers throughout the input token sequence, and 2) recycling registers for final decision predictions. We term this new architecture MambaReg. Qualitative observations suggest, compared to vanilla Vision Mamba, MambaReg's feature maps appear cleaner and more focused on semantically meaningful regions. Quantitatively, MambaReg attains stronger performance and scales better. For example, on the ImageNet benchmark, our MambaReg-B attains 83.0% accuracy, significantly outperforming Vim-B's 81.8%; furthermore, we provide the first successful scaling to the large model size (i.e., with 340M parameters), attaining a competitive accuracy of 83.6% (84.5% if finetuned with 384x384 inputs). Additional validation on the downstream semantic segmentation task also supports MambaReg's efficacy.

Currently, medical vision language models are widely used in medical vision question answering tasks. However, existing models are confronted with two issues: for input, the model only relies on text instructions and lacks direct understanding of visual clues in the image; for output, the model only gives text answers and lacks connection with key areas in the image. To address these issues, we propose a unified medical vision language model MIMO, with visual referring Multimodal Input and pixel grounding Multimodal Output. MIMO can not only combine visual clues and textual instructions to understand complex medical images and semantics, but can also ground medical terminologies in textual output within the image. To overcome the scarcity of relevant data in the medical field, we propose MIMOSeg, a comprehensive medical multimodal dataset including 895K samples. MIMOSeg is constructed from four different perspectives, covering basic instruction following and complex question answering with multimodal input and multimodal output. We conduct experiments on several downstream medical multimodal tasks. Extensive experimental results verify that MIMO can uniquely combine visual referring and pixel grounding capabilities, which are not available in previous models. Our project can be found in https://github.com/pkusixspace/MIMO.

Current structural pruning methods face two significant limitations: (i) they often limit pruning to finer-grained levels like channels, making aggressive parameter reduction challenging, and (ii) they focus heavily on parameter and FLOP reduction, with existing latency-aware methods frequently relying on simplistic, suboptimal linear models that fail to generalize well to transformers, where multiple interacting dimensions impact latency. In this paper, we address both limitations by introducing Multi-Dimensional Pruning(MDP), a novel paradigm that jointly optimizes across a variety of pruning granularities--including channels, query/key, heads, embeddings, and blocks. MDP employs an advanced latency modeling technique to accurately capture latency variations across all prunable dimensions, achieving an optimal balance between latency and accuracy. By reformulating pruning as a Mixed-Integer Nonlinear Program (MINLP), MDP efficiently identifies the optimal pruned structure across all prunable dimensions while respecting latency constraints. This versatile framework supports both CNNs and transformers. Extensive experiments demonstrate that MDP significantly outperforms previous methods, especially at high pruning ratios. On ImageNet, MDP achieves a 28% speed increase with a +1.4 Top-1 accuracy improvement over prior work like HALP for ResNet50 pruning. Against the latest transformer pruning method, Isomorphic, MDP delivers an additional 37% acceleration with a +0.7 Top-1 accuracy improvement.

Vision-Language Models (VLMs) have already enabled a variety of real-world applications. The large parameter size of VLMs brings large memory and computation overhead which poses significant challenges for deployment. Post-Training Quantization (PTQ) is an effective technique to reduce the memory and computation overhead. Existing PTQ methods mainly focus on the language modality in large language models (LLMs), without considering the differences across other modalities. In this paper, we discover that there is a significant difference in sensitivity between language and vision tokens in large VLMs. Therefore, treating tokens from different modalities equally, as in existing PTQ methods, may over-emphasize the insensitive modalities, leading to significant accuracy loss. To deal with the above issue, we propose a simple yet effective method, Modality-Balanced Quantization (MBQ), for large VLMs. Specifically, MBQ incorporates the different sensitivities across modalities during the calibration process to minimize the reconstruction loss for better quantization parameters. Extensive experiments show that MBQ can significantly improve task accuracy by up to 4.4% and 11.6% under W3A16 and W4A8 quantization for 7B to 70B VLMs, compared to SOTA baselines. Additionally, we implement a W3A16 GPU kernel that fuses the dequantization and GEMV operators, achieving a 1.4x speedup on LLaVA-onevision-7B on the RTX 4090. We will release the code.

Hybrid Mamba-Transformer networks have recently garnered broad attention. These networks can leverage the scalability of Transformers while capitalizing on Mamba's strengths in long-context modeling and computational efficiency. However, the challenge of effectively pretraining such hybrid networks remains an open question. Existing methods, such as Masked Autoencoders (MAE) or autoregressive (AR) pretraining, primarily focus on single-type network architectures. In contrast, pretraining strategies for hybrid architectures must be effective for both Mamba and Transformer components. Based on this, we propose Masked Autoregressive Pretraining (MAP) to pretrain a hybrid Mamba-Transformer vision backbone network. This strategy combines the strengths of both MAE and Autoregressive pretraining, improving the performance of Mamba and Transformer modules within a unified paradigm. Experimental results show that the hybrid Mamba-Transformer vision backbone network pretrained with MAP significantly outperforms other pretraining strategies, achieving state-of-the-art performance. We validate the method's effectiveness on both 2D and 3D datasets and provide detailed ablation studies to support the design choices for each component.