Daily Papers

Vision-language foundation models have shown remarkable performance in various zero-shot settings such as image retrieval, classification, or captioning. But so far, those models seem to fall behind when it comes to zero-shot localization of referential expressions and objects in images. As a result, they need to be fine-tuned for this task. In this paper, we show that pretrained vision-language (VL) models allow for zero-shot open-vocabulary object localization without any fine-tuning. To leverage those capabilities, we propose a Grounding Everything Module (GEM) that generalizes the idea of value-value attention introduced by CLIPSurgery to a self-self attention path. We show that the concept of self-self attention corresponds to clustering, thus enforcing groups of tokens arising from the same object to be similar while preserving the alignment with the language space. To further guide the group formation, we propose a set of regularizations that allows the model to finally generalize across datasets and backbones. We evaluate the proposed GEM framework on various benchmark tasks and datasets for semantic segmentation. It shows that GEM not only outperforms other training-free open-vocabulary localization methods, but also achieves state-of-the-art results on the recently proposed OpenImagesV7 large-scale segmentation benchmark.

In this work, we present the Textless Vision-Language Transformer (TVLT), where homogeneous transformer blocks take raw visual and audio inputs for vision-and-language representation learning with minimal modality-specific design, and do not use text-specific modules such as tokenization or automatic speech recognition (ASR). TVLT is trained by reconstructing masked patches of continuous video frames and audio spectrograms (masked autoencoding) and contrastive modeling to align video and audio. TVLT attains performance comparable to its text-based counterpart on various multimodal tasks, such as visual question answering, image retrieval, video retrieval, and multimodal sentiment analysis, with 28x faster inference speed and only 1/3 of the parameters. Our findings suggest the possibility of learning compact and efficient visual-linguistic representations from low-level visual and audio signals without assuming the prior existence of text. Our code and checkpoints are available at: https://github.com/zinengtang/TVLT

Recent vision-language models have achieved tremendous advances. However, their computational costs are also escalating dramatically, making model acceleration exceedingly critical. To pursue more efficient vision-language Transformers, this paper introduces Cross-Guided Ensemble of Tokens (CrossGET), a general acceleration framework for vision-language Transformers. This framework adaptively combines tokens in real-time during inference, significantly reducing computational costs while maintaining high performance. CrossGET features two primary innovations: 1) Cross-Guided Matching and Ensemble. CrossGET leverages cross-modal guided token matching and ensemble to effectively utilize cross-modal information, achieving wider applicability across both modality-independent models, e.g., CLIP, and modality-dependent ones, e.g., BLIP2. 2) Complete-Graph Soft Matching. CrossGET introduces an algorithm for the token-matching mechanism, ensuring reliable matching results while facilitating parallelizability and high efficiency. Extensive experiments have been conducted on various vision-language tasks, such as image-text retrieval, visual reasoning, image captioning, and visual question answering. The performance on both classic multimodal architectures and emerging multimodal LLMs demonstrates the framework's effectiveness and versatility. The code is available at https://github.com/sdc17/CrossGET.

Real-world data contains a vast amount of multimodal information, among which vision and language are the two most representative modalities. Moreover, increasingly heavier models, e.g., Transformers, have attracted the attention of researchers to model compression. However, how to compress multimodal models, especially vison-language Transformers, is still under-explored. This paper proposes the Unified and Progressive Pruning (\emph{UPop}) as a universal vison-language Transformer compression framework, which incorporates 1) unifiedly searching multimodal subnets in a continuous optimization space from the original model, which enables automatic assignment of pruning ratios among compressible modalities and structures; 2) progressively searching and retraining the subnet, which maintains convergence between the search and retrain to attain higher compression ratios. Experiments on various tasks, datasets, and model architectures demonstrate the effectiveness and versatility of the proposed UPop framework. The code is available at https://github.com/sdc17/UPop.

We present a pre-training approach for vision and language transformer models, which is based on a mixture of diverse tasks. We explore both the use of image-text captioning data in pre-training, which does not need additional supervision, as well as object-aware strategies to pre-train the model. We evaluate the method on a number of textgenerative vision+language tasks, such as Visual Question Answering, visual entailment and captioning, and demonstrate large gains over standard pre-training methods.

Vision-language models have revolutionized human-computer interaction and shown significant progress in multi-modal tasks. However, applying these models to complex visual tasks like medical image analysis remains challenging. In this study, we propose ViLaM, a unified Vision-Language transformer model that integrates instruction tuning predicated on a large language model. This approach enables us to optimally utilize the knowledge and reasoning capacities of large pre-trained language models for an array of tasks encompassing both language and vision. We employ frozen pre-trained encoders to encode and align both image and text features, enabling ViLaM to handle a variety of visual tasks following textual instructions. Besides, we've designed cycle training for referring expressions to address the need for high-quality, paired referring expression datasets for training large models in terms of both quantity and quality. We evaluated ViLaM's exceptional performance on public general datasets and further confirmed its generalizability on medical datasets. Importantly, we've observed the model's impressive zero-shot learning ability, indicating the potential future application of ViLaM in the medical field.

Current state-of-the-art vision-and-language models are evaluated on tasks either individually or in a multi-task setting, overlooking the challenges of continually learning (CL) tasks as they arrive. Existing CL benchmarks have facilitated research on task adaptation and mitigating "catastrophic forgetting", but are limited to vision-only and language-only tasks. We present CLiMB, a benchmark to study the challenge of learning multimodal tasks in a CL setting, and to systematically evaluate how upstream continual learning can rapidly generalize to new multimodal and unimodal tasks. CLiMB includes implementations of several CL algorithms and a modified Vision-Language Transformer (ViLT) model that can be deployed on both multimodal and unimodal tasks. We find that common CL methods can help mitigate forgetting during multimodal task learning, but do not enable cross-task knowledge transfer. We envision that CLiMB will facilitate research on a new class of CL algorithms for this challenging multimodal setting.

Medical phrase grounding (MPG) aims to locate the most relevant region in a medical image, given a phrase query describing certain medical findings, which is an important task for medical image analysis and radiological diagnosis. However, existing visual grounding methods rely on general visual features for identifying objects in natural images and are not capable of capturing the subtle and specialized features of medical findings, leading to sub-optimal performance in MPG. In this paper, we propose MedRPG, an end-to-end approach for MPG. MedRPG is built on a lightweight vision-language transformer encoder and directly predicts the box coordinates of mentioned medical findings, which can be trained with limited medical data, making it a valuable tool in medical image analysis. To enable MedRPG to locate nuanced medical findings with better region-phrase correspondences, we further propose Tri-attention Context contrastive alignment (TaCo). TaCo seeks context alignment to pull both the features and attention outputs of relevant region-phrase pairs close together while pushing those of irrelevant regions far away. This ensures that the final box prediction depends more on its finding-specific regions and phrases. Experimental results on three MPG datasets demonstrate that our MedRPG outperforms state-of-the-art visual grounding approaches by a large margin. Additionally, the proposed TaCo strategy is effective in enhancing finding localization ability and reducing spurious region-phrase correlations.

Applying a pre-trained large model to downstream tasks is prohibitive under resource-constrained conditions. Recent dominant approaches for addressing efficiency issues involve adding a few learnable parameters to the fixed backbone model. This strategy, however, leads to more challenges in loading large models for downstream fine-tuning with limited resources. In this paper, we propose a novel method for increasing the parameter efficiency of pre-trained models by introducing an intermediate pre-training stage. To this end, we first employ low-rank approximation to compress the original large model and then devise a feature distillation module and a weight perturbation regularization module. These modules are specifically designed to enhance the low-rank model. In particular, we update only the low-rank model while freezing the backbone parameters during pre-training. This allows for direct and efficient utilization of the low-rank model for downstream fine-tuning tasks. The proposed method achieves both efficiencies in terms of required parameters and computation time while maintaining comparable results with minimal modifications to the backbone architecture. Specifically, when applied to three vision-only and one vision-language Transformer models, our approach often demonstrates a merely sim0.6 point decrease in performance while reducing the original parameter size by 1/3 to 2/3.

Vision-and-Language Pre-training (VLP) has improved performance on various joint vision-and-language downstream tasks. Current approaches to VLP heavily rely on image feature extraction processes, most of which involve region supervision (e.g., object detection) and the convolutional architecture (e.g., ResNet). Although disregarded in the literature, we find it problematic in terms of both (1) efficiency/speed, that simply extracting input features requires much more computation than the multimodal interaction steps; and (2) expressive power, as it is upper bounded to the expressive power of the visual embedder and its predefined visual vocabulary. In this paper, we present a minimal VLP model, Vision-and-Language Transformer (ViLT), monolithic in the sense that the processing of visual inputs is drastically simplified to just the same convolution-free manner that we process textual inputs. We show that ViLT is up to tens of times faster than previous VLP models, yet with competitive or better downstream task performance. Our code and pre-trained weights are available at https://github.com/dandelin/vilt.

The increasing availability of multimodal data across text, tables, and images presents new challenges for developing models capable of complex cross-modal reasoning. Existing methods for Multimodal Multi-hop Question Answering (MMQA) often suffer from limited reasoning capabilities, reliance on modality conversion, and inadequate alignment between visual and textual representations. To address these limitations, this paper introduces Vision-Language Multimodal Transformer (VLMT), a unified architecture that integrates a transformer-based vision encoder with a sequence-to-sequence language model. VLMT employs a direct token-level injection mechanism to fuse visual and textual inputs within a shared embedding space, eliminating the need for intermediate projection layers. To enhance cross-modal alignment and reasoning, a three-stage pretraining strategy is proposed to progressively align vision-language representations and improve the model's capacity for multimodal understanding. Based on the pretrained backbone, two task-specific modules are instantiated to form a two-stage MMQA framework: a multimodal reranker that predicts document relevance scores and utilizes a relative threshold with top-k strategy for context retrieval, and a multimodal question answering model that generates contextually grounded answers based on the retrieved evidence. Comprehensive experiments on two benchmark datasets demonstrate the effectiveness of the proposed approach. On MultimodalQA validation set, VLMT-Large achieves 76.5% Exact Match and 80.1% F1, outperforming the previous state-of-the-art by +9.1% in Exact Match and +8.8% in F1. On WebQA, it attains a QA score of 47.6, surpassing prior models such as PERQA by +3.2. These results highlight VLMT's strong capabilities in multimodal reasoning and its potential to advance real-world information retrieval and question answering systems.

Cross-modal retrieval is gaining increasing efficacy and interest from the research community, thanks to large-scale training, novel architectural and learning designs, and its application in LLMs and multimodal LLMs. In this paper, we move a step forward and design an approach that allows for multimodal queries, composed of both an image and a text, and can search within collections of multimodal documents, where images and text are interleaved. Our model, ReT, employs multi-level representations extracted from different layers of both visual and textual backbones, both at the query and document side. To allow for multi-level and cross-modal understanding and feature extraction, ReT employs a novel Transformer-based recurrent cell that integrates both textual and visual features at different layers, and leverages sigmoidal gates inspired by the classical design of LSTMs. Extensive experiments on M2KR and M-BEIR benchmarks show that ReT achieves state-of-the-art performance across diverse settings. Our source code and trained models are publicly available at https://github.com/aimagelab/ReT.

Despite achieving remarkable performance on various vision-language tasks, Transformer-based Vision-Language Models (VLMs) suffer from redundancy in inputs and parameters, significantly hampering their efficiency in real-world applications. Moreover, the degree of redundancy in token representations and model parameters, such as attention heads, varies significantly for different inputs. In light of the challenges, we propose SmartTrim, an adaptive acceleration framework for VLMs, which adjusts the computational overhead per instance. Specifically, we integrate lightweight modules into the original backbone to identify and prune redundant token representations and attention heads within each layer. Furthermore, we devise a self-distillation strategy to enhance the consistency between the predictions of the pruned model and its fully-capacity counterpart. Experimental results across various vision-language tasks consistently demonstrate that SmartTrim accelerates the original model by 2-3 times with minimal performance degradation, highlighting the effectiveness and efficiency compared to previous approaches. Code will be available at https://github.com/kugwzk/SmartTrim.

We perform an effective-theory analysis of forward-backward signal propagation in wide and deep Transformers, i.e., residual neural networks with multi-head self-attention blocks and multilayer perceptron blocks. This analysis suggests particular width scalings of initialization and training hyperparameters for these models. We then take up such suggestions, training Vision and Language Transformers in practical setups.

Anti-Muslim hate speech has emerged within memes, characterized by context-dependent and rhetorical messages using text and images that seemingly mimic humor but convey Islamophobic sentiments. This work presents a novel dataset and proposes a classifier based on the Vision-and-Language Transformer (ViLT) specifically tailored to identify anti-Muslim hate within memes by integrating both visual and textual representations. Our model leverages joint modal embeddings between meme images and incorporated text to capture nuanced Islamophobic narratives that are unique to meme culture, providing both high detection accuracy and interoperability.

Large pre-trained transformers are show-stealer in modern-day deep learning, and it becomes crucial to comprehend the parsimonious patterns that exist within them as they grow in scale. With exploding parameter counts, Lottery Ticket Hypothesis (LTH) and its variants, have lost their pragmatism in sparsifying them due to high computation and memory bottleneck of repetitive train-prune-retrain routine of iterative magnitude pruning (IMP) which worsens with increasing model size. This paper comprehensively studies induced sparse patterns across multiple large pre-trained vision and language transformers. We propose the existence of -- essential sparsity defined with a sharp dropping point beyond which the performance declines much faster w.r.t the rise of sparsity level, when we directly remove weights with the smallest magnitudes in one-shot without re-training. We also find essential sparsity to hold valid for N:M sparsity patterns as well as on modern-scale large language models (Vicuna-7B). We also present an intriguing emerging phenomenon of abrupt sparsification during the pre-training of BERT, i.e., BERT suddenly becomes heavily sparse in pre-training after certain iterations. Moreover, our observations also indicate a counter-intuitive finding that BERT trained with a larger amount of pre-training data tends to have a better ability to condense knowledge in comparatively relatively fewer parameters. Lastly, we investigate the effect of the pre-training loss on essential sparsity and discover that self-supervised learning (SSL) objectives trigger stronger emergent sparsification properties than supervised learning (SL). Our codes are available at https://github.com/VITA-Group/essential_sparsity.

Referring image segmentation is a fundamental vision-language task that aims to segment out an object referred to by a natural language expression from an image. One of the key challenges behind this task is leveraging the referring expression for highlighting relevant positions in the image. A paradigm for tackling this problem is to leverage a powerful vision-language ("cross-modal") decoder to fuse features independently extracted from a vision encoder and a language encoder. Recent methods have made remarkable advancements in this paradigm by exploiting Transformers as cross-modal decoders, concurrent to the Transformer's overwhelming success in many other vision-language tasks. Adopting a different approach in this work, we show that significantly better cross-modal alignments can be achieved through the early fusion of linguistic and visual features in intermediate layers of a vision Transformer encoder network. By conducting cross-modal feature fusion in the visual feature encoding stage, we can leverage the well-proven correlation modeling power of a Transformer encoder for excavating helpful multi-modal context. This way, accurate segmentation results are readily harvested with a light-weight mask predictor. Without bells and whistles, our method surpasses the previous state-of-the-art methods on RefCOCO, RefCOCO+, and G-Ref by large margins.

We introduce TimeViper, a hybrid vision-language model designed to tackle challenges of long video understanding. Processing long videos demands both an efficient model architecture and an effective mechanism for handling extended temporal contexts. To this end, TimeViper adopts a hybrid Mamba-Transformer backbone that combines the efficiency of state-space models with the expressivity of attention mechanisms. Through this hybrid design, we reveal the vision-to-text information aggregation phenomenon, where information progressively flows from vision tokens to text tokens across increasing LLM depth, resulting in severe vision token redundancy. Motivated by this observation, we propose TransV, a token information transfer module that transfers and compresses vision tokens into instruction tokens while maintaining multimodal understanding capabilities. This design enables TimeViper to process hour-long videos exceeding 10,000 frames. Extensive experiments across multiple benchmarks demonstrate that TimeViper competes with state-of-the-art models while extending frame numbers. We further analyze attention behaviors of both Mamba and Transformer layers, offering new insights into hybrid model interpretability. This work represents an initial step towards developing, interpreting, and compressing hybrid Mamba-Transformer architectures.

In this work, we explore neat yet effective Transformer-based frameworks for visual grounding. The previous methods generally address the core problem of visual grounding, i.e., multi-modal fusion and reasoning, with manually-designed mechanisms. Such heuristic designs are not only complicated but also make models easily overfit specific data distributions. To avoid this, we first propose TransVG, which establishes multi-modal correspondences by Transformers and localizes referred regions by directly regressing box coordinates. We empirically show that complicated fusion modules can be replaced by a simple stack of Transformer encoder layers with higher performance. However, the core fusion Transformer in TransVG is stand-alone against uni-modal encoders, and thus should be trained from scratch on limited visual grounding data, which makes it hard to be optimized and leads to sub-optimal performance. To this end, we further introduce TransVG++ to make two-fold improvements. For one thing, we upgrade our framework to a purely Transformer-based one by leveraging Vision Transformer (ViT) for vision feature encoding. For another, we devise Language Conditioned Vision Transformer that removes external fusion modules and reuses the uni-modal ViT for vision-language fusion at the intermediate layers. We conduct extensive experiments on five prevalent datasets, and report a series of state-of-the-art records.

This paper reveals that large language models (LLMs), despite being trained solely on textual data, are surprisingly strong encoders for purely visual tasks in the absence of language. Even more intriguingly, this can be achieved by a simple yet previously overlooked strategy -- employing a frozen transformer block from pre-trained LLMs as a constituent encoder layer to directly process visual tokens. Our work pushes the boundaries of leveraging LLMs for computer vision tasks, significantly departing from conventional practices that typically necessitate a multi-modal vision-language setup with associated language prompts, inputs, or outputs. We demonstrate that our approach consistently enhances performance across a diverse range of tasks, encompassing pure 2D and 3D visual recognition tasks (e.g., image and point cloud classification), temporal modeling tasks (e.g., action recognition), non-semantic tasks (e.g., motion forecasting), and multi-modal tasks (e.g., 2D/3D visual question answering and image-text retrieval). Such improvements are a general phenomenon, applicable to various types of LLMs (e.g., LLaMA and OPT) and different LLM transformer blocks. We additionally propose the information filtering hypothesis to explain the effectiveness of pre-trained LLMs in visual encoding -- the pre-trained LLM transformer blocks discern informative visual tokens and further amplify their effect. This hypothesis is empirically supported by the observation that the feature activation, after training with LLM transformer blocks, exhibits a stronger focus on relevant regions. We hope that our work inspires new perspectives on utilizing LLMs and deepening our understanding of their underlying mechanisms. Code is available at https://github.com/ziqipang/LM4VisualEncoding.

Integrating visual-language instructions into visuomotor policies is gaining momentum in robot learning for enhancing open-world generalization. Despite promising advances, existing approaches face two challenges: limited language steerability when no generated reasoning is used as a condition, or significant inference latency when reasoning is incorporated. In this work, we introduce MoTVLA, a mixture-of-transformers (MoT)-based vision-language-action (VLA) model that integrates fast-slow unified reasoning with behavior policy learning. MoTVLA preserves the general intelligence of pre-trained VLMs (serving as the generalist) for tasks such as perception, scene understanding, and semantic planning, while incorporating a domain expert, a second transformer that shares knowledge with the pretrained VLM, to generate domain-specific fast reasoning (e.g., robot motion decomposition), thereby improving policy execution efficiency. By conditioning the action expert on decomposed motion instructions, MoTVLA can learn diverse behaviors and substantially improve language steerability. Extensive evaluations across natural language processing benchmarks, robotic simulation environments, and real-world experiments confirm the superiority of MoTVLA in both fast-slow reasoning and manipulation task performance.

Contrastive Language-Image Pre-training (CLIP) has significantly improved performance in various vision-language tasks by expanding the dataset with image-text pairs obtained from websites. This paper further explores CLIP from the perspectives of data and model architecture. To address the prevalence of noisy data and enhance the quality of large-scale image-text data crawled from the internet, we introduce a diverse description generation framework that can leverage Large Language Models (LLMs) to synthesize and refine content from web-based texts, synthetic captions, and detection tags. Furthermore, we propose RWKV-CLIP, the first RWKV-driven vision-language representation learning model that combines the effective parallel training of transformers with the efficient inference of RNNs. Comprehensive experiments across various model scales and pre-training datasets demonstrate that RWKV-CLIP is a robust and efficient vision-language representation learner, it achieves state-of-the-art performance in several downstream tasks, including linear probe, zero-shot classification, and zero-shot image-text retrieval. To facilitate future research, the code and pre-trained models are released at https://github.com/deepglint/RWKV-CLIP

The Large Vision-Language Model (LVLM) integrates computer vision and natural language processing techniques, offering substantial application potential. However, these models demand extensive resources during inference. Adaptive attention techniques can dynamically reduce computational redundancy and thus improve efficiency. Although current adaptive attention methods significantly reduce the memory requirements of Transformer-based language models, they are not tailored for LVLMs. We observe that LVLMs generate responses from both remote image tokens and local text tokens, and different modalities have different attention patterns. This observation inspires us to manage the attention for each modality separately. Specifically, for visual input, we store the cache of potentially useful information but only compute the most critical parts. For language input, we care more about local information. Based on our observation and analysis of vision-language attention patterns, we develop A-VL, a plug-and-play adaptive attention tailored for LVLM inference. Extensive evaluations on three vision-language tasks and five datasets show the effectiveness of our designs. Our approach A-VL outperforms existing adaptive attention methods in reducing memory usage and computational load without compromising performance.

Despite the remarkable success of foundation models, their task-specific fine-tuning paradigm makes them inconsistent with the goal of general perception modeling. The key to eliminating this inconsistency is to use generalist models for general task modeling. However, existing attempts at generalist models are inadequate in both versatility and performance. In this paper, we propose Uni-Perceiver v2, which is the first generalist model capable of handling major large-scale vision and vision-language tasks with competitive performance. Specifically, images are encoded as general region proposals, while texts are encoded via a Transformer-based language model. The encoded representations are transformed by a task-agnostic decoder. Different tasks are formulated as a unified maximum likelihood estimation problem. We further propose an improved optimizer to ensure stable multi-task learning with an unmixed sampling strategy, which is helpful for tasks requiring large batch-size training. After being jointly trained on various tasks, Uni-Perceiver v2 is capable of directly handling downstream tasks without any task-specific adaptation. Results show that Uni-Perceiver v2 outperforms all existing generalist models in both versatility and performance. Meanwhile, compared with the commonly-recognized strong baselines that require tasks-specific fine-tuning, Uni-Perceiver v2 achieves competitive performance on a broad range of vision and vision-language tasks.

Measuring how real images look is a complex task in artificial intelligence research. For example, an image of a boy with a vacuum cleaner in a desert violates common sense. We introduce a novel method, which we call Through the Looking Glass (TLG), to assess image common sense consistency using Large Vision-Language Models (LVLMs) and Transformer-based encoder. By leveraging LVLMs to extract atomic facts from these images, we obtain a mix of accurate facts. We proceed by fine-tuning a compact attention-pooling classifier over encoded atomic facts. Our TLG has achieved a new state-of-the-art performance on the WHOOPS! and WEIRD datasets while leveraging a compact fine-tuning component.

Large Vision-Language Models (LVLMs) typically encode an image into a fixed number of visual tokens (e.g., 576) and process these tokens with a language model. Despite their strong performance, LVLMs face challenges in adapting to varying computational constraints. This raises the question: can we achieve flexibility in the number of visual tokens to suit different tasks and computational resources? We answer this with an emphatic yes. Inspired by Matryoshka Representation Learning, we introduce the Matryoshka Query Transformer (MQT), capable of encoding an image into m visual tokens during inference, where m can be any number up to a predefined maximum. This is achieved by employing a query transformer with M latent query tokens to compress the visual embeddings. During each training step, we randomly select m <= M latent query tokens and train the model using only these first m tokens, discarding the rest. Combining MQT with LLaVA, we train a single model once, and flexibly and drastically reduce the number of inference-time visual tokens while maintaining similar or better performance compared to training independent models for each number of tokens. Our model, MQT-LLAVA, matches LLaVA-1.5 performance across 11 benchmarks using a maximum of 256 tokens instead of LLaVA's fixed 576. Reducing to 16 tokens (8x less TFLOPs) only sacrifices the performance by 2.4 points on MMBench. On certain tasks such as ScienceQA and MMMU, we can even go down to only 2 visual tokens with performance drops of just 3% and 6% each. Our exploration of the trade-off between the accuracy and computational cost brought about by the number of visual tokens facilitates future research to achieve the best of both worlds.

This work tackles the information loss bottleneck of vector-quantization (VQ) autoregressive image generation by introducing a novel model architecture called the 2-Dimensional Autoregression (DnD) Transformer. The DnD-Transformer predicts more codes for an image by introducing a new autoregression direction, model depth, along with the sequence length direction. Compared to traditional 1D autoregression and previous work utilizing similar 2D image decomposition such as RQ-Transformer, the DnD-Transformer is an end-to-end model that can generate higher quality images with the same backbone model size and sequence length, opening a new optimization perspective for autoregressive image generation. Furthermore, our experiments reveal that the DnD-Transformer's potential extends beyond generating natural images. It can even generate images with rich text and graphical elements in a self-supervised manner, demonstrating an understanding of these combined modalities. This has not been previously demonstrated for popular vision generative models such as diffusion models, showing a spark of vision-language intelligence when trained solely on images. Code, datasets and models are open at https://github.com/chenllliang/DnD-Transformer.

We present SOLO, a single transformer for Scalable visiOn-Language mOdeling. Current large vision-language models (LVLMs) such as LLaVA mostly employ heterogeneous architectures that connect pre-trained visual encoders with large language models (LLMs) to facilitate visual recognition and complex reasoning. Although achieving remarkable performance with relatively lightweight training, we identify four primary scalability limitations: (1) The visual capacity is constrained by pre-trained visual encoders, which are typically an order of magnitude smaller than LLMs. (2) The heterogeneous architecture complicates the use of established hardware and software infrastructure. (3) Study of scaling laws on such architecture must consider three separate components - visual encoder, connector, and LLMs, which complicates the analysis. (4) The use of existing visual encoders typically requires following a pre-defined specification of image inputs pre-processing, for example, by reshaping inputs to fixed-resolution square images, which presents difficulties in processing and training on high-resolution images or those with unusual aspect ratio. A unified single Transformer architecture, like SOLO, effectively addresses these scalability concerns in LVLMs; however, its limited adoption in the modern context likely stems from the absence of reliable training recipes that balance both modalities and ensure stable training for billion-scale models. In this paper, we introduce the first open-source training recipe for developing SOLO, an open-source 7B LVLM using moderate academic resources. The training recipe involves initializing from LLMs, sequential pre-training on ImageNet and web-scale data, and instruction fine-tuning on our curated high-quality datasets. On extensive evaluation, SOLO demonstrates performance comparable to LLaVA-v1.5-7B, particularly excelling in visual mathematical reasoning.

While recent vision-language-action models trained on diverse robot datasets exhibit promising generalization capabilities with limited in-domain data, their reliance on compact action heads to predict discretized or continuous actions constrains adaptability to heterogeneous action spaces. We present Dita, a scalable framework that leverages Transformer architectures to directly denoise continuous action sequences through a unified multimodal diffusion process. Departing from prior methods that condition denoising on fused embeddings via shallow networks, Dita employs in-context conditioning -- enabling fine-grained alignment between denoised actions and raw visual tokens from historical observations. This design explicitly models action deltas and environmental nuances. By scaling the diffusion action denoiser alongside the Transformer's scalability, Dita effectively integrates cross-embodiment datasets across diverse camera perspectives, observation scenes, tasks, and action spaces. Such synergy enhances robustness against various variances and facilitates the successful execution of long-horizon tasks. Evaluations across extensive benchmarks demonstrate state-of-the-art or comparative performance in simulation. Notably, Dita achieves robust real-world adaptation to environmental variances and complex long-horizon tasks through 10-shot finetuning, using only third-person camera inputs. The architecture establishes a versatile, lightweight and open-source baseline for generalist robot policy learning. Project Page: https://robodita.github.io.

Successful generalist Vision-Language-Action (VLA) models rely on effective training across diverse robotic platforms with large-scale, cross-embodiment, heterogeneous datasets. To facilitate and leverage the heterogeneity in rich, diverse robotic data sources, we propose a novel Soft Prompt approach with minimally added parameters, by infusing prompt learning concepts into cross-embodiment robot learning and introducing separate sets of learnable embeddings for each distinct data source. These embeddings serve as embodiment-specific prompts, which in unity empower VLA models with effective exploitation of varying cross-embodiment features. Our new X-VLA, a neat flow-matching-based VLA architecture, relies exclusively on soft-prompted standard Transformer encoders, enjoying both scalability and simplicity. Evaluated across 6 simulations as well as 3 real-world robots, our 0.9B instantiation-X-VLA-0.9B simultaneously achieves SOTA performance over a sweep of benchmarks, demonstrating superior results on a wide axes of capabilities, from flexible dexterity to quick adaptation across embodiments, environments, and tasks. Website: https://thu-air-dream.github.io/X-VLA/

With the advancement of RNN models with linear complexity, the quadratic complexity challenge of transformers has the potential to be overcome. Notably, the emerging Mamba-2 has demonstrated competitive performance, bridging the gap between RNN models and transformers. However, due to sequential processing and vanishing gradients, RNN models struggle to capture long-range dependencies, limiting contextual understanding. This results in slow convergence, high resource demands, and poor performance on downstream understanding and complex reasoning tasks. In this work, we present a hybrid model MaTVLM by substituting a portion of the transformer decoder layers in a pre-trained VLM with Mamba-2 layers. Leveraging the inherent relationship between attention and Mamba-2, we initialize Mamba-2 with corresponding attention weights to accelerate convergence. Subsequently, we employ a single-stage distillation process, using the pre-trained VLM as the teacher model to transfer knowledge to the MaTVLM, further enhancing convergence speed and performance. Furthermore, we investigate the impact of differential distillation loss within our training framework. We evaluate the MaTVLM on multiple benchmarks, demonstrating competitive performance against the teacher model and existing VLMs while surpassing both Mamba-based VLMs and models of comparable parameter scales. Remarkably, the MaTVLM achieves up to 3.6x faster inference than the teacher model while reducing GPU memory consumption by 27.5%, all without compromising performance. Code and models are released at http://github.com/hustvl/MaTVLM.

Multi-head self-attention (MHSA) is a key component of Transformers, a widely popular architecture in both language and vision. Multiple heads intuitively enable different parallel processes over the same input. Yet, they also obscure the attribution of each input patch to the output of a model. We propose a novel 1-head Transformer Attention Bottleneck (TAB) layer, inserted after the traditional MHSA architecture, to serve as an attention bottleneck for interpretability and intervention. Unlike standard self-attention, TAB constrains the total attention over all patches to in [0, 1]. That is, when the total attention is 0, no visual information is propagated further into the network and the vision-language model (VLM) would default to a generic, image-independent response. To demonstrate the advantages of TAB, we train VLMs with TAB to perform image difference captioning. Over three datasets, our models perform similarly to baseline VLMs in captioning but the bottleneck is superior in localizing changes and in identifying when no changes occur. TAB is the first architecture to enable users to intervene by editing attention, which often produces expected outputs by VLMs.

This study explores replacing Transformers in Visual Language Models (VLMs) with Mamba, a recent structured state space model (SSM) that demonstrates promising performance in sequence modeling. We test models up to 3B parameters under controlled conditions, showing that Mamba-based VLMs outperforms Transformers-based VLMs in captioning, question answering, and reading comprehension. However, we find that Transformers achieve greater performance in visual grounding and the performance gap widens with scale. We explore two hypotheses to explain this phenomenon: 1) the effect of task-agnostic visual encoding on the updates of the hidden states, and 2) the difficulty in performing visual grounding from the perspective of in-context multimodal retrieval. Our results indicate that a task-aware encoding yields minimal performance gains on grounding, however, Transformers significantly outperform Mamba at in-context multimodal retrieval. Overall, Mamba shows promising performance on tasks where the correct output relies on a summary of the image but struggles when retrieval of explicit information from the context is required.

This paper introduces SAIL, a single transformer unified multimodal large language model (MLLM) that integrates raw pixel encoding and language decoding within a singular architecture. Unlike existing modular MLLMs, which rely on a pre-trained vision transformer (ViT), SAIL eliminates the need for a separate vision encoder, presenting a more minimalist architecture design. Instead of introducing novel architectural components, SAIL adapts mix-attention mechanisms and multimodal positional encodings to better align with the distinct characteristics of visual and textual modalities. We systematically compare SAIL's properties-including scalability, cross-modal information flow patterns, and visual representation capabilities-with those of modular MLLMs. By scaling both training data and model size, SAIL achieves performance comparable to modular MLLMs. Notably, the removal of pretrained ViT components enhances SAIL's scalability and results in significantly different cross-modal information flow patterns. Moreover, SAIL demonstrates strong visual representation capabilities, achieving results on par with ViT-22B in vision tasks such as semantic segmentation. Code and models are available at https://github.com/bytedance/SAIL.

In this work, we introduce Vision-Language Generative Pre-trained Transformer (VL-GPT), a transformer model proficient at concurrently perceiving and generating visual and linguistic data. VL-GPT achieves a unified pre-training approach for both image and text modalities by employing a straightforward auto-regressive objective, thereby enabling the model to process image and text as seamlessly as a language model processes text. To accomplish this, we initially propose a novel image tokenizer-detokenizer framework for visual data, specifically designed to transform raw images into a sequence of continuous embeddings and reconstruct them accordingly. In combination with the existing text tokenizer and detokenizer, this framework allows for the encoding of interleaved image-text data into a multimodal sequence, which can subsequently be fed into the transformer model. Consequently, VL-GPT can perform large-scale pre-training on multimodal corpora utilizing a unified auto-regressive objective (i.e., next-token prediction). Upon completion of pre-training, VL-GPT exhibits remarkable zero-shot and few-shot performance across a diverse range of vision and language understanding and generation tasks, including image captioning, visual question answering, text-to-image generation, and more. Additionally, the pre-trained model retrains in-context learning capabilities when provided with multimodal prompts. We further conduct instruction tuning on our VL-GPT, highlighting its exceptional potential for multimodal assistance. The source code and model weights shall be released.

In this paper, we design and train a Generative Image-to-text Transformer, GIT, to unify vision-language tasks such as image/video captioning and question answering. While generative models provide a consistent network architecture between pre-training and fine-tuning, existing work typically contains complex structures (uni/multi-modal encoder/decoder) and depends on external modules such as object detectors/taggers and optical character recognition (OCR). In GIT, we simplify the architecture as one image encoder and one text decoder under a single language modeling task. We also scale up the pre-training data and the model size to boost the model performance. Without bells and whistles, our GIT establishes new state of the arts on 12 challenging benchmarks with a large margin. For instance, our model surpasses the human performance for the first time on TextCaps (138.2 vs. 125.5 in CIDEr). Furthermore, we present a new scheme of generation-based image classification and scene text recognition, achieving decent performance on standard benchmarks. Codes are released at https://github.com/microsoft/GenerativeImage2Text.

Vision-Language (VL) models have gained significant research focus, enabling remarkable advances in multimodal reasoning. These architectures typically comprise a vision encoder, a Large Language Model (LLM), and a projection module that aligns visual features with the LLM's representation space. Despite their success, a critical limitation persists: the vision encoding process remains decoupled from user queries, often in the form of image-related questions. Consequently, the resulting visual features may not be optimally attuned to the query-specific elements of the image. To address this, we introduce QA-ViT, a Question Aware Vision Transformer approach for multimodal reasoning, which embeds question awareness directly within the vision encoder. This integration results in dynamic visual features focusing on relevant image aspects to the posed question. QA-ViT is model-agnostic and can be incorporated efficiently into any VL architecture. Extensive experiments demonstrate the effectiveness of applying our method to various multimodal architectures, leading to consistent improvement across diverse tasks and showcasing its potential for enhancing visual and scene-text understanding.

Radiology plays a pivotal role in modern medicine due to its non-invasive diagnostic capabilities. However, the manual generation of unstructured medical reports is time consuming and prone to errors. It creates a significant bottleneck in clinical workflows. Despite advancements in AI-generated radiology reports, challenges remain in achieving detailed and accurate report generation. In this study we have evaluated different combinations of multimodal models that integrate Computer Vision and Natural Language Processing to generate comprehensive radiology reports. We employed a pretrained Vision Transformer (ViT-B16) and a SWIN Transformer as the image encoders. The BART and GPT-2 models serve as the textual decoders. We used Chest X-ray images and reports from the IU-Xray dataset to evaluate the usability of the SWIN Transformer-BART, SWIN Transformer-GPT-2, ViT-B16-BART and ViT-B16-GPT-2 models for report generation. We aimed at finding the best combination among the models. The SWIN-BART model performs as the best-performing model among the four models achieving remarkable results in almost all the evaluation metrics like ROUGE, BLEU and BERTScore.

Open-vocabulary dense prediction tasks including object detection and image segmentation have been advanced by the success of Contrastive Language-Image Pre-training (CLIP). CLIP models, particularly those incorporating vision transformers (ViTs), have exhibited remarkable generalization ability in zero-shot image classification. However, when transferring the vision-language alignment of CLIP from global image representation to local region representation for the open-vocabulary dense prediction tasks, CLIP ViTs suffer from the domain shift from full images to local image regions. In this paper, we embark on an in-depth analysis of the region-language alignment in CLIP models, which is essential for downstream open-vocabulary dense prediction tasks. Subsequently, we propose an approach named CLIPSelf, which adapts the image-level recognition ability of CLIP ViT to local image regions without needing any region-text pairs. CLIPSelf empowers ViTs to distill itself by aligning a region representation extracted from its dense feature map with the image-level representation of the corresponding image crop. With the enhanced CLIP ViTs, we achieve new state-of-the-art performance on open-vocabulary object detection, semantic segmentation, and panoptic segmentation across various benchmarks. Models and code will be available at https://github.com/wusize/CLIPSelf.

Vision Transformer (ViT) based Vision-Language Pre-training (VLP) models have demonstrated impressive performance in various tasks. However, the lengthy visual token sequences fed into ViT can lead to training inefficiency and ineffectiveness. Existing efforts address the challenge by either bottom-level patch extraction in the ViT backbone or top-level patch abstraction outside, not balancing training efficiency and effectiveness well. Inspired by text summarization in natural language processing, we propose a Bottom-Up Patch Summarization approach named BUS, coordinating bottom-level extraction and top-level abstraction to learn a concise summary of lengthy visual token sequences efficiently. Specifically, We incorporate a Text-Semantics-Aware Patch Selector (TSPS) into the ViT backbone to perform a coarse-grained visual token extraction and then attach a flexible Transformer-based Patch Abstraction Decoder (PAD) upon the backbone for top-level visual abstraction. This bottom-up collaboration enables our BUS to yield high training efficiency while maintaining or even improving effectiveness. We evaluate our approach on various visual-language understanding and generation tasks and show competitive downstream task performance while boosting the training efficiency by 50\%. Additionally, our model achieves state-of-the-art performance on many downstream tasks by increasing input image resolution without increasing computational costs over baselines.

Contrastive pre-training on image-text pairs, exemplified by CLIP, becomes a standard technique for learning multi-modal visual-language representations. Although CLIP has demonstrated remarkable performance, training it from scratch on noisy web-scale datasets is computationally demanding. On the other hand, mask-then-predict pre-training approaches, like Masked Image Modeling (MIM), offer efficient self-supervised learning for single-modal representations. This paper introduces Unmasked Token Alignment (UTA), a method that leverages existing CLIP models to further enhance its vision-language representations. UTA trains a Vision Transformer (ViT) by aligning unmasked visual tokens to the corresponding image tokens from a frozen CLIP vision encoder, which automatically aligns the ViT model with the CLIP text encoder. The pre-trained ViT can be directly applied for zero-shot evaluation even without training on image-text pairs. Compared to MIM approaches, UTA does not suffer from training-finetuning inconsistency and is much more training-efficient by avoiding using the extra [MASK] tokens. Extensive experimental results demonstrate that UTA can enhance CLIP models and outperform existing MIM methods on various uni- and multi-modal benchmarks. Code and models are available at https://github.com/jihaonew/UTA.

Recent advancements in vision-language models (VLMs) have expanded their potential for real-world applications, enabling these models to perform complex reasoning on images. In the widely used fully autoregressive transformer-based models like LLaVA, projected visual tokens are prepended to textual tokens. Oftentimes, visual tokens are significantly more than prompt tokens, resulting in increased computational overhead during both training and inference. In this paper, we propose Visual Compact Token Registers (Victor), a method that reduces the number of visual tokens by summarizing them into a smaller set of register tokens. Victor adds a few learnable register tokens after the visual tokens and summarizes the visual information into these registers using the first few layers in the language tower of VLMs. After these few layers, all visual tokens are discarded, significantly improving computational efficiency for both training and inference. Notably, our method is easy to implement and requires a small number of new trainable parameters with minimal impact on model performance. In our experiment, with merely 8 visual registers--about 1% of the original tokens--Victor shows less than a 4% accuracy drop while reducing the total training time by 43% and boosting the inference throughput by 3.3X.

3D vision-language grounding (3D-VL) is an emerging field that aims to connect the 3D physical world with natural language, which is crucial for achieving embodied intelligence. Current 3D-VL models rely heavily on sophisticated modules, auxiliary losses, and optimization tricks, which calls for a simple and unified model. In this paper, we propose 3D-VisTA, a pre-trained Transformer for 3D Vision and Text Alignment that can be easily adapted to various downstream tasks. 3D-VisTA simply utilizes self-attention layers for both single-modal modeling and multi-modal fusion without any sophisticated task-specific design. To further enhance its performance on 3D-VL tasks, we construct ScanScribe, the first large-scale 3D scene-text pairs dataset for 3D-VL pre-training. ScanScribe contains 2,995 RGB-D scans for 1,185 unique indoor scenes originating from ScanNet and 3R-Scan datasets, along with paired 278K scene descriptions generated from existing 3D-VL tasks, templates, and GPT-3. 3D-VisTA is pre-trained on ScanScribe via masked language/object modeling and scene-text matching. It achieves state-of-the-art results on various 3D-VL tasks, ranging from visual grounding and dense captioning to question answering and situated reasoning. Moreover, 3D-VisTA demonstrates superior data efficiency, obtaining strong performance even with limited annotations during downstream task fine-tuning.

We present a novel methodology aimed at optimizing the application of frozen large language models (LLMs) for resource-intensive vision-language (VL) pre-training. The current paradigm uses visual features as prompts to guide language models, with a focus on determining the most relevant visual features for corresponding text. Our approach diverges by concentrating on the language component, specifically identifying the optimal prompts to align with visual features. We introduce the Prompt-Transformer (P-Former), a model that predicts these ideal prompts, which is trained exclusively on linguistic data, bypassing the need for image-text pairings. This strategy subtly bifurcates the end-to-end VL training process into an additional, separate stage. Our experiments reveal that our framework significantly enhances the performance of a robust image-to-text baseline (BLIP-2), and effectively narrows the performance gap between models trained with either 4M or 129M image-text pairs. Importantly, our framework is modality-agnostic and flexible in terms of architectural design, as validated by its successful application in a video learning task using varied base modules. The code is available at https://github.com/yiren-jian/BLIText

Recent breakthroughs in vision-language models (VLMs) start a new page in the vision community. The VLMs provide stronger and more generalizable feature embeddings compared to those from ImageNet-pretrained models, thanks to the training on the large-scale Internet image-text pairs. However, despite the amazing achievement from the VLMs, vanilla Vision Transformers (ViTs) remain the default choice for the image encoder. Although pure transformer proves its effectiveness in the text encoding area, it remains questionable whether it is also the case for image encoding, especially considering that various types of networks are proposed on the ImageNet benchmark, which, unfortunately, are rarely studied in VLMs. Due to small data/model scale, the original conclusions of model design on ImageNet can be limited and biased. In this paper, we aim at building an evaluation protocol of vision models in the vision-language era under the contrastive language-image pretraining (CLIP) framework. We provide a comprehensive way to benchmark different vision models, covering their zero-shot performance and scalability in both model and training data sizes. To this end, we introduce ViTamin, a new vision models tailored for VLMs. ViTamin-L significantly outperforms ViT-L by 2.0% ImageNet zero-shot accuracy, when using the same publicly available DataComp-1B dataset and the same OpenCLIP training scheme. ViTamin-L presents promising results on 60 diverse benchmarks, including classification, retrieval, open-vocabulary detection and segmentation, and large multi-modal models. When further scaling up the model size, our ViTamin-XL with only 436M parameters attains 82.9% ImageNet zero-shot accuracy, surpassing 82.0% achieved by EVA-E that has ten times more parameters (4.4B).

This paper presents ViTOC (Vision Transformer and Object-aware Captioner), a novel vision-language model for image captioning that addresses the challenges of accuracy and diversity in generated descriptions. Unlike conventional approaches, ViTOC employs a dual-path architecture based on Vision Transformer and object detector, effectively fusing global visual features and local object information through learnable vectors. The model introduces an innovative object-aware prompting strategy that significantly enhances its capability in handling long-tail data. Experiments on the standard COCO dataset demonstrate that ViTOC outperforms baseline models across all evaluation metrics. Additionally, we propose a reference-free evaluation method based on CLIP to further validate the model's effectiveness. By utilizing pretrained visual model parameters, ViTOC achieves efficient end-to-end training.

We investigate the mechanism underlying a previously identified phenomenon in Vision Transformers -- the emergence of high-norm tokens that lead to noisy attention maps. We observe that in multiple models (e.g., CLIP, DINOv2), a sparse set of neurons is responsible for concentrating high-norm activations on outlier tokens, leading to irregular attention patterns and degrading downstream visual processing. While the existing solution for removing these outliers involves retraining models from scratch with additional learned register tokens, we use our findings to create a training-free approach to mitigate these artifacts. By shifting the high-norm activations from our discovered register neurons into an additional untrained token, we can mimic the effect of register tokens on a model already trained without registers. We demonstrate that our method produces cleaner attention and feature maps, enhances performance over base models across multiple downstream visual tasks, and achieves results comparable to models explicitly trained with register tokens. We then extend test-time registers to off-the-shelf vision-language models to improve their interpretability. Our results suggest that test-time registers effectively take on the role of register tokens at test-time, offering a training-free solution for any pre-trained model released without them.

With Transformers achieving outstanding performance on individual remote sensing (RS) tasks, we are now approaching the realization of a unified model that excels across multiple tasks through multi-task learning (MTL). Compared to single-task approaches, MTL methods offer improved generalization, enhanced scalability, and greater practical applicability. Recently, vision language models (VLMs) have achieved promising results in RS image understanding, grounding, and ultra-high-resolution (UHR) image reasoning, respectively. Moreover, the unified text-based interface demonstrates significant potential for MTL. Hence, in this work, we present RSCoVLM, a simple yet flexible VLM baseline for RS MTL. Firstly, we create the data curation engine, including data acquisition, offline processing and integrating, as well as online loading and weighting. This data engine effectively addresses complex RS data enviroment and generates flexible vision-language conversations. Furthermore, we propose a unified dynamic-resolution strategy to address the diverse image scales inherent in RS imagery. For UHR images, we introduce the Zoom-in Chain mechanism together with its corresponding dataset, LRS-VQA-Zoom. The strategies are flexible and effectively mitigate the computational burdens. Additionally, we significantly enhance the model's object detection capability and propose a novel evaluation protocol that ensures fair comparison between VLMs and conventional detection models. Extensive experiments demonstrate that RSCoVLM achieves state-of-the-art performance across diverse tasks, outperforming existing RS VLMs and even rivaling specialized expert models. All the training and evaluating tools, model weights, and datasets have been fully open-sourced to support reproducibility. We expect that this baseline will promote further progress toward general-purpose RS models.

Executing language-conditioned tasks in dynamic visual environments remains a central challenge in embodied AI. Existing Vision-Language-Action (VLA) models predominantly adopt reactive state-to-action mappings, often leading to short-sighted behaviors and poor robustness in dynamic scenes. In this paper, we introduce F1, a pretrained VLA framework which integrates the visual foresight generation into decision-making pipeline. F1 adopts a Mixture-of-Transformer architecture with dedicated modules for perception, foresight generation, and control, thereby bridging understanding, generation, and actions. At its core, F1 employs a next-scale prediction mechanism to synthesize goal-conditioned visual foresight as explicit planning targets. By forecasting plausible future visual states, F1 reformulates action generation as a foresight-guided inverse dynamics problem, enabling actions that implicitly achieve visual goals. To endow F1 with robust and generalizable capabilities, we propose a three-stage training recipe on an extensive dataset comprising over 330k trajectories across 136 diverse tasks. This training scheme enhances modular reasoning and equips the model with transferable visual foresight, which is critical for complex and dynamic environments. Extensive evaluations on real-world tasks and simulation benchmarks demonstrate F1 consistently outperforms existing approaches, achieving substantial gains in both task success rate and generalization ability.

Vision-Language-Action (VLA) models are a promising path to realizing generalist embodied agents that can quickly adapt to new tasks, modalities, and environments. However, methods for interpreting and steering VLAs fall far short of classical robotics pipelines, which are grounded in explicit models of kinematics, dynamics, and control. This lack of mechanistic insight is a central challenge for deploying learned policies in real-world robotics, where robustness and explainability are critical. Motivated by advances in mechanistic interpretability for large language models, we introduce the first framework for interpreting and steering VLAs via their internal representations, enabling direct intervention in model behavior at inference time. We project feedforward activations within transformer layers onto the token embedding basis, identifying sparse semantic directions - such as speed and direction - that are causally linked to action selection. Leveraging these findings, we introduce a general-purpose activation steering method that modulates behavior in real time, without fine-tuning, reward signals, or environment interaction. We evaluate this method on two recent open-source VLAs, Pi0 and OpenVLA, and demonstrate zero-shot behavioral control in simulation (LIBERO) and on a physical robot (UR5). This work demonstrates that interpretable components of embodied VLAs can be systematically harnessed for control - establishing a new paradigm for transparent and steerable foundation models in robotics.

Self-supervised learning in vision-language processing exploits semantic alignment between imaging and text modalities. Prior work in biomedical VLP has mostly relied on the alignment of single image and report pairs even though clinical notes commonly refer to prior images. This does not only introduce poor alignment between the modalities but also a missed opportunity to exploit rich self-supervision through existing temporal content in the data. In this work, we explicitly account for prior images and reports when available during both training and fine-tuning. Our approach, named BioViL-T, uses a CNN-Transformer hybrid multi-image encoder trained jointly with a text model. It is designed to be versatile to arising challenges such as pose variations and missing input images across time. The resulting model excels on downstream tasks both in single- and multi-image setups, achieving state-of-the-art performance on (I) progression classification, (II) phrase grounding, and (III) report generation, whilst offering consistent improvements on disease classification and sentence-similarity tasks. We release a novel multi-modal temporal benchmark dataset, MS-CXR-T, to quantify the quality of vision-language representations in terms of temporal semantics. Our experimental results show the advantages of incorporating prior images and reports to make most use of the data.

This paper presents a unified Vision-Language Pre-training (VLP) model. The model is unified in that (1) it can be fine-tuned for either vision-language generation (e.g., image captioning) or understanding (e.g., visual question answering) tasks, and (2) it uses a shared multi-layer transformer network for both encoding and decoding, which differs from many existing methods where the encoder and decoder are implemented using separate models. The unified VLP model is pre-trained on a large amount of image-text pairs using the unsupervised learning objectives of two tasks: bidirectional and sequence-to-sequence (seq2seq) masked vision-language prediction. The two tasks differ solely in what context the prediction conditions on. This is controlled by utilizing specific self-attention masks for the shared transformer network. To the best of our knowledge, VLP is the first reported model that achieves state-of-the-art results on both vision-language generation and understanding tasks, as disparate as image captioning and visual question answering, across three challenging benchmark datasets: COCO Captions, Flickr30k Captions, and VQA 2.0. The code and the pre-trained models are available at https://github.com/LuoweiZhou/VLP.

Transformer-based models have driven significant advancements in Multimodal Large Language Models (MLLMs), yet their computational costs surge drastically when scaling resolution, training data, and model parameters. A key bottleneck stems from the proliferation of visual tokens required for fine-grained image understanding. We propose Skip-Vision, a unified framework addressing both training and inference inefficiencies in vision-language models. On top of conventional token compression approaches, our method introduces two complementary acceleration strategies. For training acceleration, we observe that Feed-Forward Network (FFN) computations on visual tokens induce marginal feature updates. This motivates our Skip-FFN strategy, which bypasses FFN layers for redundant visual tokens. For inference acceleration, we design a selective KV-cache removal mechanism that prunes the skipped key-value pairs during decoding while preserving model performance. Experimental results demonstrate that Skip-Vision reduces training time by up to 35\%, inference FLOPs by 75\%, and latency by 45\%, while achieving comparable or superior performance to existing methods. Our work provides a practical solution for scaling high-performance MLLMs with enhanced efficiency.

Vision-language (VL) pre-training has recently received considerable attention. However, most existing end-to-end pre-training approaches either only aim to tackle VL tasks such as image-text retrieval, visual question answering (VQA) and image captioning that test high-level understanding of images, or only target region-level understanding for tasks such as phrase grounding and object detection. We present FIBER (Fusion-In-the-Backbone-based transformER), a new VL model architecture that can seamlessly handle both these types of tasks. Instead of having dedicated transformer layers for fusion after the uni-modal backbones, FIBER pushes multimodal fusion deep into the model by inserting cross-attention into the image and text backbones, bringing gains in terms of memory and performance. In addition, unlike previous work that is either only pre-trained on image-text data or on fine-grained data with box-level annotations, we present a two-stage pre-training strategy that uses both these kinds of data efficiently: (i) coarse-grained pre-training based on image-text data; followed by (ii) fine-grained pre-training based on image-text-box data. We conduct comprehensive experiments on a wide range of VL tasks, ranging from VQA, image captioning, and retrieval, to phrase grounding, referring expression comprehension, and object detection. Using deep multimodal fusion coupled with the two-stage pre-training, FIBER provides consistent performance improvements over strong baselines across all tasks, often outperforming methods using magnitudes more data. Code is available at https://github.com/microsoft/FIBER.

Vision-Language-Action (VLA) models have recently shown impressive generalization and language-guided manipulation capabilities. However, their performance degrades on tasks requiring precise spatial reasoning due to limited spatial reasoning inherited from Vision-Language Models (VLMs). Existing VLAs rely on extensive action-data pretraining to ground VLMs in 3D space, which reduces training efficiency and is still insufficient for accurate spatial understanding. In this work, we present DepthVLA, a simple yet effective VLA architecture that explicitly incorporates spatial awareness through a pretrained depth prediction module. DepthVLA adopts a mixture-of-transformers design that unifies a VLM, a depth transformer, and an action expert with fully shared attentions, forming an end-to-end model with enhanced spatial reasoning. Extensive evaluations in both real-world and simulated environments show that DepthVLA outperforms state-of-the-art approaches, achieving 78.5% vs. 65.0% progress in real-world tasks, 94.9% vs. 93.6% in the LIBERO simulator, and 74.8% vs. 58.8% in the Simpler simulator. Our code will be made publicly available.

Temporal action detection (TAD) involves the localization and classification of action instances within untrimmed videos. While standard TAD follows fully supervised learning with closed-set setting on large training data, recent zero-shot TAD methods showcase the promising open-set setting by leveraging large-scale contrastive visual-language (ViL) pretrained models. However, existing zero-shot TAD methods have limitations on how to properly construct the strong relationship between two interdependent tasks of localization and classification and adapt ViL model to video understanding. In this work, we present ZEETAD, featuring two modules: dual-localization and zero-shot proposal classification. The former is a Transformer-based module that detects action events while selectively collecting crucial semantic embeddings for later recognition. The latter one, CLIP-based module, generates semantic embeddings from text and frame inputs for each temporal unit. Additionally, we enhance discriminative capability on unseen classes by minimally updating the frozen CLIP encoder with lightweight adapters. Extensive experiments on THUMOS14 and ActivityNet-1.3 datasets demonstrate our approach's superior performance in zero-shot TAD and effective knowledge transfer from ViL models to unseen action categories.

Vision language models (VLMs) have seen growing adoption in recent years, but many still struggle with basic spatial reasoning errors. We hypothesize that this is due to VLMs adopting pre-trained vision backbones, specifically vision transformers (ViTs) trained with image-level supervision and minimal inductive biases. Such models may fail to encode the class contents at each position in the image, and our goal is to resolve this by ensuring that the vision backbone effectively captures both local and global image semantics. Our main insight is that we do not require new supervision to learn this capability -- pre-trained models contain significant knowledge of local semantics that we can extract and use for scalable self-supervision. We propose a new efficient post-training stage for ViTs called locality alignment and a novel fine-tuning procedure called MaskEmbed that uses a masked reconstruction loss to learn semantic contributions for each image patch. We first evaluate locality alignment with a vision-only benchmark, finding that it improves a model's performance at a patch-level semantic segmentation task, especially for strong backbones trained with image-caption pairs (e.g., CLIP and SigLIP). We then train a series of VLMs with and without locality alignment, and show that locality-aligned backbones improve performance across a range of benchmarks, particularly ones that involve spatial understanding (e.g., RefCOCO, OCID-Ref, TallyQA, VSR, AI2D). Overall, we demonstrate that we can efficiently learn local semantic extraction via a locality alignment stage, and that this procedure complements existing VLM training recipes that use off-the-shelf vision backbones.

We present Florence-VL, a new family of multimodal large language models (MLLMs) with enriched visual representations produced by Florence-2, a generative vision foundation model. Unlike the widely used CLIP-style vision transformer trained by contrastive learning, Florence-2 can capture different levels and aspects of visual features, which are more versatile to be adapted to diverse downstream tasks. We propose a novel feature-fusion architecture and an innovative training recipe that effectively integrates Florence-2's visual features into pretrained LLMs, such as Phi 3.5 and LLama 3. In particular, we propose "depth-breath fusion (DBFusion)" to fuse the visual features extracted from different depths and under multiple prompts. Our model training is composed of end-to-end pretraining of the whole model followed by finetuning of the projection layer and the LLM, on a carefully designed recipe of diverse open-source datasets that include high-quality image captions and instruction-tuning pairs. Our quantitative analysis and visualization of Florence-VL's visual features show its advantages over popular vision encoders on vision-language alignment, where the enriched depth and breath play important roles. Florence-VL achieves significant improvements over existing state-of-the-art MLLMs across various multi-modal and vision-centric benchmarks covering general VQA, perception, hallucination, OCR, Chart, knowledge-intensive understanding, etc. To facilitate future research, our models and the complete training recipe are open-sourced. https://github.com/JiuhaiChen/Florence-VL

Although recent advances in visual generation have been remarkable, most existing architectures still depend on distinct encoders for images and text. This separation constrains diffusion models' ability to perform cross-modal reasoning and knowledge transfer. Prior attempts to bridge this gap often use the last layer information from VLM, employ multiple visual encoders, or train large unified models jointly for text and image generation, which demands substantial computational resources and large-scale data, limiting its accessibility.We present UniFusion, a diffusion-based generative model conditioned on a frozen large vision-language model (VLM) that serves as a unified multimodal encoder. At the core of UniFusion is the Layerwise Attention Pooling (LAP) mechanism that extracts both high level semantics and low level details from text and visual tokens of a frozen VLM to condition a diffusion generative model. We demonstrate that LAP outperforms other shallow fusion architectures on text-image alignment for generation and faithful transfer of visual information from VLM to the diffusion model which is key for editing. We propose VLM-Enabled Rewriting Injection with Flexibile Inference (VERIFI), which conditions a diffusion transformer (DiT) only on the text tokens generated by the VLM during in-model prompt rewriting. VERIFI combines the alignment of the conditioning distribution with the VLM's reasoning capabilities for increased capabilities and flexibility at inference. In addition, finetuning on editing task not only improves text-image alignment for generation, indicative of cross-modality knowledge transfer, but also exhibits tremendous generalization capabilities. Our model when trained on single image editing, zero-shot generalizes to multiple image references further motivating the unified encoder design of UniFusion.

Recently, augmenting Vision-Language-Action models (VLAs) with world modeling has shown promise in improving robotic policy learning. However, it remains challenging to jointly predict next-state observations and action sequences because of the inherent difference between the two modalities. To address this, we propose DUal-STream diffusion (DUST), a world-model augmented VLA framework that handles the modality conflict and enhances the performance of VLAs across diverse tasks. Specifically, we propose a multimodal diffusion transformer architecture that explicitly maintains separate modality streams while still enabling cross-modal knowledge sharing. In addition, we introduce independent noise perturbations for each modality and a decoupled flow-matching loss. This design enables the model to learn the joint distribution in a bidirectional manner while avoiding the need for a unified latent space. Based on the decoupling of modalities during training, we also introduce a joint sampling method that supports test-time scaling, where action and vision tokens evolve asynchronously at different rates. Through experiments on simulated benchmarks such as RoboCasa and GR-1, DUST achieves up to 6% gains over baseline methods, while our test-time scaling approach provides an additional 2-5% boost. On real-world tasks with the Franka Research 3, DUST improves success rates by 13%, confirming its effectiveness beyond simulation. Furthermore, pre-training on action-free videos from BridgeV2 yields significant transfer gains on RoboCasa, underscoring DUST's potential for large-scale VLA pretraining.

We present Cephalo, a series of multimodal vision large language models (V-LLMs) designed for materials science applications, integrating visual and linguistic data for enhanced understanding and interaction within human-AI and multi-agent AI frameworks. A key innovation of Cephalo is its advanced dataset generation method, which employs a sophisticated algorithm to accurately detect and separate images and their corresponding textual descriptions from PDF documents, such as scientific papers. The method includes a careful refinement of image-text pairs through integrated vision and language processing, ensuring high-quality, contextually relevant, and well reasoned training data. Cephalo is trained on integrated image and text data extracted from thousands of scientific papers and science-focused Wikipedia pages demonstrates can interpret complex visual scenes, generate precise language descriptions, and answer queries about images effectively. The combination of a vision encoder with an autoregressive transformer supports complex natural language understanding in an integrated model, which can be coupled with other generative methods to create an image-to-text-to-image or image-to-text-to-3D pipeline. To explore the development of larger models from smaller ones, we merge sets of layers that originate from different pre-trained source models. This hybrid approach allows us to leverage the domain-specific expertise and general conversational capabilities to harness the strengths of multiple models. We examine the models in diverse use cases that incorporate biological materials, fracture and engineering analysis, protein biophysics, and bio-inspired design based on insect behavior. Generative applications include bio-inspired designs, including pollen-inspired architected materials, as well as the synthesis of bio-inspired material microstructures from a photograph of a solar eclipse.

Vision-language pre-training for chest X-rays has made significant strides, primarily by utilizing paired radiographs and radiology reports. However, existing approaches often face challenges in encoding medical knowledge effectively. While radiology reports provide insights into the current disease manifestation, medical definitions (as used by contemporary methods) tend to be overly abstract, creating a gap in knowledge. To address this, we propose DeViDe, a novel transformer-based method that leverages radiographic descriptions from the open web. These descriptions outline general visual characteristics of diseases in radiographs, and when combined with abstract definitions and radiology reports, provide a holistic snapshot of knowledge. DeViDe incorporates three key features for knowledge-augmented vision language alignment: First, a large-language model-based augmentation is employed to homogenise medical knowledge from diverse sources. Second, this knowledge is aligned with image information at various levels of granularity. Third, a novel projection layer is proposed to handle the complexity of aligning each image with multiple descriptions arising in a multi-label setting. In zero-shot settings, DeViDe performs comparably to fully supervised models on external datasets and achieves state-of-the-art results on three large-scale datasets. Additionally, fine-tuning DeViDe on four downstream tasks and six segmentation tasks showcases its superior performance across data from diverse distributions.

Vision-language pre-training (VLP) methods are blossoming recently, and its crucial goal is to jointly learn visual and textual features via a transformer-based architecture, demonstrating promising improvements on a variety of vision-language tasks. Prior arts usually focus on how to align visual and textual features, but strategies for improving the robustness of model and speeding up model convergence are left insufficiently explored. In this paper, we propose a novel method ViLTA, comprising of two components to further facilitate the model to learn fine-grained representations among image-text pairs. For Masked Language Modeling (MLM), we propose a cross-distillation method to generate soft labels to enhance the robustness of model, which alleviates the problem of treating synonyms of masked words as negative samples in one-hot labels. For Image-Text Matching (ITM), we leverage the current language encoder to synthesize hard negatives based on the context of language input, encouraging the model to learn high-quality representations by increasing the difficulty of the ITM task. By leveraging the above techniques, our ViLTA can achieve better performance on various vision-language tasks. Extensive experiments on benchmark datasets demonstrate that the effectiveness of ViLTA and its promising potential for vision-language pre-training.

The advancement of large Vision-Language-Action (VLA) models has significantly improved robotic manipulation in terms of language-guided task execution and generalization to unseen scenarios. While existing VLAs adapted from pretrained large Vision-Language-Models (VLM) have demonstrated promising generalizability, their task performance is still unsatisfactory as indicated by the low tasks success rates in different environments. In this paper, we present a new advanced VLA architecture derived from VLM. Unlike previous works that directly repurpose VLM for action prediction by simple action quantization, we propose a omponentized VLA architecture that has a specialized action module conditioned on VLM output. We systematically study the design of the action module and demonstrates the strong performance enhancement with diffusion action transformers for action sequence modeling, as well as their favorable scaling behaviors. We also conduct comprehensive experiments and ablation studies to evaluate the efficacy of our models with varied designs. The evaluation on 5 robot embodiments in simulation and real work shows that our model not only significantly surpasses existing VLAs in task performance and but also exhibits remarkable adaptation to new robots and generalization to unseen objects and backgrounds. It exceeds the average success rates of OpenVLA which has similar model size (7B) with ours by over 35% in simulated evaluation and 55% in real robot experiments. It also outperforms the large RT-2-X model (55B) by 18% absolute success rates in simulation. Code and models can be found on our project page (https://cogact.github.io/).

When trained at a sufficient scale, self-supervised learning has exhibited a notable ability to solve a wide range of visual or language understanding tasks. In this paper, we investigate simple, yet effective approaches for adapting the pre-trained foundation models to the downstream task of interest, namely, open-vocabulary semantic segmentation. To this end, we make the following contributions: (i) we introduce Fusioner, with a lightweight, transformer-based fusion module, that pairs the frozen visual representation with language concept through a handful of image segmentation data. As a consequence, the model gains the capability of zero-shot transfer to segment novel categories; (ii) without loss of generality, we experiment on a broad range of self-supervised models that have been pre-trained with different schemes, e.g. visual-only models (MoCo v3, DINO), language-only models (BERT), visual-language model (CLIP), and show that, the proposed fusion approach is effective to any pair of visual and language models, even those pre-trained on a corpus of uni-modal data; (iii) we conduct thorough ablation studies to analyze the critical components in our proposed Fusioner, while evaluating on standard benchmarks, e.g. PASCAL-5i and COCO-20i , it surpasses existing state-of-the-art models by a large margin, despite only being trained on frozen visual and language features; (iv) to measure the model's robustness on learning visual-language correspondence, we further evaluate on synthetic dataset, named Mosaic-4, where images are constructed by mosaicking the samples from FSS-1000. Fusioner demonstrates superior performance over previous models.

We present OmniJARVIS, a novel Vision-Language-Action (VLA) model for open-world instruction-following agents in open-world Minecraft. Compared to prior works that either emit textual goals to separate controllers or produce the control command directly, OmniJARVIS seeks a different path to ensure both strong reasoning and efficient decision-making capabilities via unified tokenization of multimodal interaction data. First, we introduce a self-supervised approach to learn a behavior encoder that produces discretized tokens for behavior trajectories tau = {o_0, a_0, dots} and an imitation learning (IL) policy decoder conditioned on these tokens. These additional behavior tokens will be augmented to the vocabulary of pretrained Multimodal Language Models (MLMs). With this encoder, we then pack long-term multimodal interactions involving task instructions, memories, thoughts, observations, textual responses, behavior trajectories, etc. into unified token sequences and model them with autoregressive transformers. Thanks to the semantically meaningful behavior tokens, the resulting VLA model, OmniJARVIS, can reason (by producing chain-of-thoughts), plan, answer questions, and act (by producing behavior tokens for the IL policy decoder). OmniJARVIS demonstrates excellent performances on a comprehensive collection of atomic, programmatic, and open-ended tasks in open-world Minecraft. Our analysis further unveils the crucial design principles in interaction data formation, unified tokenization, and its scaling potentials.

The advent of next-generation radio telescopes is set to transform radio astronomy by producing massive data volumes that challenge traditional processing methods. Deep learning techniques have shown strong potential in automating radio analysis tasks, yet are often constrained by the limited availability of large annotated datasets. Recent progress in self-supervised learning has led to foundational radio vision models, but adapting them for new tasks typically requires coding expertise, limiting their accessibility to a broader astronomical community. Text-based AI interfaces offer a promising alternative by enabling task-specific queries and example-driven learning. In this context, Large Language Models (LLMs), with their remarkable zero-shot capabilities, are increasingly used in scientific domains. However, deploying large-scale models remains resource-intensive, and there is a growing demand for AI systems that can reason over both visual and textual data in astronomical analysis. This study explores small-scale Vision-Language Models (VLMs) as AI assistants for radio astronomy, combining LLM capabilities with vision transformers. We fine-tuned the LLaVA VLM on a dataset of 59k radio images from multiple surveys, enriched with 38k image-caption pairs from the literature. The fine-tuned models show clear improvements over base models in radio-specific tasks, achieving ~30% F1-score gains in extended source detection, but they underperform pure vision models and exhibit ~20% drop on general multimodal tasks. Inclusion of caption data and LoRA fine-tuning enhances instruction-following and helps recover ~10% accuracy on standard benchmarks. This work lays the foundation for future advancements in radio VLMs, highlighting their potential and limitations, such as the need for better multimodal alignment, higher-quality datasets, and mitigation of catastrophic forgetting.

Large Pre-trained Transformers exhibit an intriguing capacity for in-context learning. Without gradient updates, these models can rapidly construct new predictors from demonstrations presented in the inputs. Recent works promote this ability in the vision-language domain by incorporating visual information into large language models that can already make in-context predictions. However, these methods could inherit issues in the language domain, such as template sensitivity and hallucination. Also, the scale of these language models raises a significant demand for computations, making learning and operating these models resource-intensive. To this end, we raise a question: ``How can we enable in-context learning without relying on the intrinsic in-context ability of large language models?". To answer it, we propose a succinct and general framework, Self-supervised IN-Context learning (SINC), that introduces a meta-model to learn on self-supervised prompts consisting of tailored demonstrations. The learned models can be transferred to downstream tasks for making in-context predictions on-the-fly. Extensive experiments show that SINC outperforms gradient-based methods in various vision-language tasks under few-shot settings. Furthermore, the designs of SINC help us investigate the benefits of in-context learning across different tasks, and the analysis further reveals the essential components for the emergence of in-context learning in the vision-language domain.

A big convergence of language, vision, and multimodal pretraining is emerging. In this work, we introduce a general-purpose multimodal foundation model BEiT-3, which achieves state-of-the-art transfer performance on both vision and vision-language tasks. Specifically, we advance the big convergence from three aspects: backbone architecture, pretraining task, and model scaling up. We introduce Multiway Transformers for general-purpose modeling, where the modular architecture enables both deep fusion and modality-specific encoding. Based on the shared backbone, we perform masked "language" modeling on images (Imglish), texts (English), and image-text pairs ("parallel sentences") in a unified manner. Experimental results show that BEiT-3 obtains state-of-the-art performance on object detection (COCO), semantic segmentation (ADE20K), image classification (ImageNet), visual reasoning (NLVR2), visual question answering (VQAv2), image captioning (COCO), and cross-modal retrieval (Flickr30K, COCO).

Despite their proficiency in general tasks, Multi-modal Large Language Models (MLLMs) struggle with automatic Geometry Problem Solving (GPS), which demands understanding diagrams, interpreting symbols, and performing complex reasoning. This limitation arises from their pre-training on natural images and texts, along with the lack of automated verification in the problem-solving process. Besides, current geometric specialists are limited by their task-specific designs, making them less effective for broader geometric problems. To this end, we present GeoX, a multi-modal large model focusing on geometric understanding and reasoning tasks. Given the significant differences between geometric diagram-symbol and natural image-text, we introduce unimodal pre-training to develop a diagram encoder and symbol decoder, enhancing the understanding of geometric images and corpora. Furthermore, we introduce geometry-language alignment, an effective pre-training paradigm that bridges the modality gap between unimodal geometric experts. We propose a Generator-And-Sampler Transformer (GS-Former) to generate discriminative queries and eliminate uninformative representations from unevenly distributed geometric signals. Finally, GeoX benefits from visual instruction tuning, empowering it to take geometric images and questions as input and generate verifiable solutions. Experiments show that GeoX outperforms both generalists and geometric specialists on publicly recognized benchmarks, such as GeoQA, UniGeo, Geometry3K, and PGPS9k.

Technologies for recognizing facial attributes like race, gender, age, and emotion have several applications, such as surveillance, advertising content, sentiment analysis, and the study of demographic trends and social behaviors. Analyzing demographic characteristics based on images and analyzing facial expressions have several challenges due to the complexity of humans' facial attributes. Traditional approaches have employed CNNs and various other deep learning techniques, trained on extensive collections of labeled images. While these methods demonstrated effective performance, there remains potential for further enhancements. In this paper, we propose to utilize vision language models (VLMs) such as generative pre-trained transformer (GPT), GEMINI, large language and vision assistant (LLAVA), PaliGemma, and Microsoft Florence2 to recognize facial attributes such as race, gender, age, and emotion from images with human faces. Various datasets like FairFace, AffectNet, and UTKFace have been utilized to evaluate the solutions. The results show that VLMs are competitive if not superior to traditional techniques. Additionally, we propose "FaceScanPaliGemma"--a fine-tuned PaliGemma model--for race, gender, age, and emotion recognition. The results show an accuracy of 81.1%, 95.8%, 80%, and 59.4% for race, gender, age group, and emotion classification, respectively, outperforming pre-trained version of PaliGemma, other VLMs, and SotA methods. Finally, we propose "FaceScanGPT", which is a GPT-4o model to recognize the above attributes when several individuals are present in the image using a prompt engineered for a person with specific facial and/or physical attributes. The results underscore the superior multitasking capability of FaceScanGPT to detect the individual's attributes like hair cut, clothing color, postures, etc., using only a prompt to drive the detection and recognition tasks.

Building scalable vision-language models to learn from diverse, multimodal data remains an open challenge. In this paper, we introduce an Efficient Vision-languagE foundation model, namely EVE, which is one unified multimodal Transformer pre-trained solely by one unified pre-training task. Specifically, EVE encodes both vision and language within a shared Transformer network integrated with modality-aware sparse Mixture-of-Experts (MoE) modules, which capture modality-specific information by selectively switching to different experts. To unify pre-training tasks of vision and language, EVE performs masked signal modeling on image-text pairs to reconstruct masked signals, i.e., image pixels and text tokens, given visible signals. This simple yet effective pre-training objective accelerates training by 3.5x compared to the model pre-trained with Image-Text Contrastive and Image-Text Matching losses. Owing to the combination of the unified architecture and pre-training task, EVE is easy to scale up, enabling better downstream performance with fewer resources and faster training speed. Despite its simplicity, EVE achieves state-of-the-art performance on various vision-language downstream tasks, including visual question answering, visual reasoning, and image-text retrieval.

The important manifestation of robot intelligence is the ability to naturally interact and autonomously make decisions. Traditional approaches to robot control often compartmentalize perception, planning, and decision-making, simplifying system design but limiting the synergy between different information streams. This compartmentalization poses challenges in achieving seamless autonomous reasoning, decision-making, and action execution. To address these limitations, a novel paradigm, named Vision-Language-Action tasks for QUAdruped Robots (QUAR-VLA), has been introduced in this paper. This approach tightly integrates visual information and instructions to generate executable actions, effectively merging perception, planning, and decision-making. The central idea is to elevate the overall intelligence of the robot. Within this framework, a notable challenge lies in aligning fine-grained instructions with visual perception information. This emphasizes the complexity involved in ensuring that the robot accurately interprets and acts upon detailed instructions in harmony with its visual observations. Consequently, we propose QUAdruped Robotic Transformer (QUART), a family of VLA models to integrate visual information and instructions from diverse modalities as input and generates executable actions for real-world robots and present QUAdruped Robot Dataset (QUARD), a large-scale multi-task dataset including navigation, complex terrain locomotion, and whole-body manipulation tasks for training QUART models. Our extensive evaluation (4000 evaluation trials) shows that our approach leads to performant robotic policies and enables QUART to obtain a range of emergent capabilities.

We present a unified Vision-Language pretrained Model (VLMo) that jointly learns a dual encoder and a fusion encoder with a modular Transformer network. Specifically, we introduce Mixture-of-Modality-Experts (MoME) Transformer, where each block contains a pool of modality-specific experts and a shared self-attention layer. Because of the modeling flexibility of MoME, pretrained VLMo can be fine-tuned as a fusion encoder for vision-language classification tasks, or used as a dual encoder for efficient image-text retrieval. Moreover, we propose a stagewise pre-training strategy, which effectively leverages large-scale image-only and text-only data besides image-text pairs. Experimental results show that VLMo achieves state-of-the-art results on various vision-language tasks, including VQA, NLVR2 and image-text retrieval. The code and pretrained models are available at https://aka.ms/vlmo.

The emergence of small vision-language models (sVLMs) marks a critical advancement in multimodal AI, enabling efficient processing of visual and textual data in resource-constrained environments. This survey offers a comprehensive exploration of sVLM development, presenting a taxonomy of architectures - transformer-based, mamba-based, and hybrid - that highlight innovations in compact design and computational efficiency. Techniques such as knowledge distillation, lightweight attention mechanisms, and modality pre-fusion are discussed as enablers of high performance with reduced resource requirements. Through an in-depth analysis of models like TinyGPT-V, MiniGPT-4, and VL-Mamba, we identify trade-offs between accuracy, efficiency, and scalability. Persistent challenges, including data biases and generalization to complex tasks, are critically examined, with proposed pathways for addressing them. By consolidating advancements in sVLMs, this work underscores their transformative potential for accessible AI, setting a foundation for future research into efficient multimodal systems.

Large-scale Transformer models bring significant improvements for various downstream vision language tasks with a unified architecture. The performance improvements come with increasing model size, resulting in slow inference speed and increased cost for severing. While some certain predictions benefit from the full complexity of the large-scale model, not all of inputs need the same amount of computation to conduct, potentially leading to computation resource waste. To handle this challenge, early exiting is proposed to adaptively allocate computational power in term of input complexity to improve inference efficiency. The existing early exiting strategies usually adopt output confidence based on intermediate layers as a proxy of input complexity to incur the decision of skipping following layers. However, such strategies cannot apply to encoder in the widely-used unified architecture with both encoder and decoder due to difficulty of output confidence estimation in the encoder. It is suboptimal in term of saving computation power to ignore the early exiting in encoder component. To handle this challenge, we propose a novel early exiting strategy for unified visual language models, which allows dynamically skip the layers in encoder and decoder simultaneously in term of input layer-wise similarities with multiple times of early exiting, namely MuE. By decomposing the image and text modalities in the encoder, MuE is flexible and can skip different layers in term of modalities, advancing the inference efficiency while minimizing performance drop. Experiments on the SNLI-VE and MS COCO datasets show that the proposed approach MuE can reduce expected inference time by up to 50\% and 40\% while maintaining 99\% and 96\% performance respectively.

We introduce a vision-language foundation model called VL-BEiT, which is a bidirectional multimodal Transformer learned by generative pretraining. Our minimalist solution conducts masked prediction on both monomodal and multimodal data with a shared Transformer. Specifically, we perform masked vision-language modeling on image-text pairs, masked language modeling on texts, and masked image modeling on images. VL-BEiT is learned from scratch with one unified pretraining task, one shared backbone, and one-stage training. Our method is conceptually simple and empirically effective. Experimental results show that VL-BEiT obtains strong results on various vision-language benchmarks, such as visual question answering, visual reasoning, and image-text retrieval. Moreover, our method learns transferable visual features, achieving competitive performance on image classification, and semantic segmentation.

Vision-Language Navigation (VLN) is a challenging task that requires an embodied agent to perform action-level modality alignment, i.e., make instruction-asked actions sequentially in complex visual environments. Most existing VLN agents learn the instruction-path data directly and cannot sufficiently explore action-level alignment knowledge inside the multi-modal inputs. In this paper, we propose modAlity-aligneD Action PrompTs (ADAPT), which provides the VLN agent with action prompts to enable the explicit learning of action-level modality alignment to pursue successful navigation. Specifically, an action prompt is defined as a modality-aligned pair of an image sub-prompt and a text sub-prompt, where the former is a single-view observation and the latter is a phrase like ''walk past the chair''. When starting navigation, the instruction-related action prompt set is retrieved from a pre-built action prompt base and passed through a prompt encoder to obtain the prompt feature. Then the prompt feature is concatenated with the original instruction feature and fed to a multi-layer transformer for action prediction. To collect high-quality action prompts into the prompt base, we use the Contrastive Language-Image Pretraining (CLIP) model which has powerful cross-modality alignment ability. A modality alignment loss and a sequential consistency loss are further introduced to enhance the alignment of the action prompt and enforce the agent to focus on the related prompt sequentially. Experimental results on both R2R and RxR show the superiority of ADAPT over state-of-the-art methods.

The success of Vision Language Models (VLMs) on various vision-language tasks heavily relies on pre-training with large scale web-crawled datasets. However, the noisy and incomplete nature of web data makes dataset scale crucial for performance, rendering end-to-end training increasingly prohibitive. In this paper, we propose NEVLP, a noise-robust framework for efficient vision-language pre-training that requires less pre-training data. Specifically, we bridge the modality gap between a frozen image encoder and a large language model with a transformer and introduce two innovative learning strategies: noise-adaptive learning and concept-enhanced learning to mitigate the impact of noise. In noise-adaptive learning, we estimate the noise probability of each image-text pair based on the transformer's memorization effect and employ noise-adaptive regularization on image-text contrastive learning to condition cross-modal alignment. In concept-enhanced learning, we enrich incomplete text by incorporating visual concepts (objects in the image) to provide prior information about existing objects for image-text matching and image-grounded text generation, thereby mitigating text incompletion. Our framework effectively utilizes noisy web data and achieves state-of-the-art performance with less pre-training data across a wide range of vision-language tasks, including image-text retrieval, image captioning, and visual question answering.

We use (multi)modal deep neural networks (DNNs) to probe for sites of multimodal integration in the human brain by predicting stereoencephalography (SEEG) recordings taken while human subjects watched movies. We operationalize sites of multimodal integration as regions where a multimodal vision-language model predicts recordings better than unimodal language, unimodal vision, or linearly-integrated language-vision models. Our target DNN models span different architectures (e.g., convolutional networks and transformers) and multimodal training techniques (e.g., cross-attention and contrastive learning). As a key enabling step, we first demonstrate that trained vision and language models systematically outperform their randomly initialized counterparts in their ability to predict SEEG signals. We then compare unimodal and multimodal models against one another. Because our target DNN models often have different architectures, number of parameters, and training sets (possibly obscuring those differences attributable to integration), we carry out a controlled comparison of two models (SLIP and SimCLR), which keep all of these attributes the same aside from input modality. Using this approach, we identify a sizable number of neural sites (on average 141 out of 1090 total sites or 12.94%) and brain regions where multimodal integration seems to occur. Additionally, we find that among the variants of multimodal training techniques we assess, CLIP-style training is the best suited for downstream prediction of the neural activity in these sites.

Multimodal Large Language Models (MLLMs) have recently sparked significant interest, which demonstrates emergent capabilities to serve as a general-purpose model for various vision-language tasks. However, existing methods mainly focus on limited types of instructions with a single image as visual context, which hinders the widespread availability of MLLMs. In this paper, we introduce the I4 benchmark to comprehensively evaluate the instruction following ability on complicated interleaved vision-language instructions, which involve intricate image-text sequential context, covering a diverse range of scenarios (e.g., visually-rich webpages/textbooks, lecture slides, embodied dialogue). Systematic evaluation on our I4 benchmark reveals a common defect of existing methods: the Visual Prompt Generator (VPG) trained on image-captioning alignment objective tends to attend to common foreground information for captioning but struggles to extract specific information required by particular tasks. To address this issue, we propose a generic and lightweight controllable knowledge re-injection module, which utilizes the sophisticated reasoning ability of LLMs to control the VPG to conditionally extract instruction-specific visual information and re-inject it into the LLM. Further, we introduce an annotation-free cross-attention guided counterfactual image training strategy to methodically learn the proposed module by collaborating a cascade of foundation models. Enhanced by the proposed module and training strategy, we present Cheetor, a Transformer-based MLLM that can effectively handle a wide variety of interleaved vision-language instructions and achieves state-of-the-art zero-shot performance across all tasks of I4, without high-quality multimodal instruction tuning data. Cheetor also exhibits competitive performance compared with state-of-the-art instruction tuned models on MME benchmark.

We study the task of extending the large language model (LLM) into a vision-language instruction-following model. This task is crucial but challenging since the LLM is trained on text modality only, making it hard to effectively digest the visual modality. To address this, existing methods typically train a visual adapter to align the representation between a pre-trained vision transformer (ViT) and the LLM by a generative image captioning loss. However, we find that the generative objective can only produce weak alignment for vision and language, making the aligned vision-language model very hungry for the instruction fine-tuning data. In this paper, we propose CG-VLM that applies both Contrastive and Generative alignment objectives to effectively align the representation of ViT and LLM. Different from image level and sentence level alignment in common contrastive learning settings, CG-VLM aligns the image-patch level features and text-token level embeddings, which, however, is very hard to achieve as no explicit grounding patch-token relation provided in standard image captioning datasets. To address this issue, we propose to maximize the averaged similarity between pooled image-patch features and text-token embeddings. Extensive experiments demonstrate that the proposed CG-VLM produces strong vision-language alignment and is an efficient instruction learner. For example, using only 10% instruction tuning data, we reach 95% performance of state-of-the-art method LLaVA [29] on the zero-shot ScienceQA-Image benchmark.

Large Language Models (LLMs) have introduced a new era of proficiency in comprehending complex healthcare and biomedical topics. However, there is a noticeable lack of models in languages other than English and models that can interpret multi-modal input, which is crucial for global healthcare accessibility. In response, this study introduces Qilin-Med-VL, the first Chinese large vision-language model designed to integrate the analysis of textual and visual data. Qilin-Med-VL combines a pre-trained Vision Transformer (ViT) with a foundational LLM. It undergoes a thorough two-stage curriculum training process that includes feature alignment and instruction tuning. This method enhances the model's ability to generate medical captions and answer complex medical queries. We also release ChiMed-VL, a dataset consisting of more than 1M image-text pairs. This dataset has been carefully curated to enable detailed and comprehensive interpretation of medical data using various types of images.

Recent advances in vision-language models have significantly expanded the frontiers of automated image analysis. However, applying these models in safety-critical contexts remains challenging due to the complex relationships between objects, subtle visual cues, and the heightened demand for transparency and reliability. This paper presents the Multi-Modal Explainable Learning (MMEL) framework, designed to enhance the interpretability of vision-language models while maintaining high performance. Building upon prior work in gradient-based explanations for transformer architectures (Grad-eclip), MMEL introduces a novel Hierarchical Semantic Relationship Module that enhances model interpretability through multi-scale feature processing, adaptive attention weighting, and cross-modal alignment. Our approach processes features at multiple semantic levels to capture relationships between image regions at different granularities, applying learnable layer-specific weights to balance contributions across the model's depth. This results in more comprehensive visual explanations that highlight both primary objects and their contextual relationships with improved precision. Through extensive experiments on standard datasets, we demonstrate that by incorporating semantic relationship information into gradient-based attribution maps, MMEL produces more focused and contextually aware visualizations that better reflect how vision-language models process complex scenes. The MMEL framework generalizes across various domains, offering valuable insights into model decisions for applications requiring high interpretability and reliability.

Self-supervised learning can resolve numerous image or linguistic processing problems when effectively trained. This study investigated simple yet efficient methods for adapting previously learned foundation models for open-vocabulary semantic segmentation tasks. Our research proposed "Beyond-Labels," a lightweight transformer-based fusion module that uses a handful of image segmentation data to fuse frozen image representations with language concepts. This strategy allows the model to successfully actualize enormous knowledge from pretrained models without requiring extensive retraining, making the model data-efficient and scalable. Furthermore, we efficiently captured positional information in images using Fourier embeddings, thus improving the generalization across various image sizes, addressing one of the key limitations of previous methods. Extensive ablation tests were performed to investigate the important components of our proposed method; when tested against the common benchmark PASCAL-5i, it demonstrated superior performance despite being trained on frozen image and language characteristics.

While Neural Processing Units (NPUs) offer high theoretical efficiency for edge AI, state-of-the-art Vision--Language Models (VLMs) tailored for GPUs often falter on these substrates. We attribute this hardware-model mismatch to two primary factors: the quantization brittleness of Vision Transformers (ViTs) and the I/O-bound nature of autoregressive attention mechanisms, which fail to utilize the high arithmetic throughput of NPUs. To bridge this gap, we propose AutoNeural, an NPU-native VLM architecture co-designed for integer-only inference. We replace the standard ViT encoder with a MobileNetV5-style backbone utilizing depthwise separable convolutions, which ensures bounded activation distributions for stable INT4/8/16 quantization. Complementing this, our language backbone integrates State-Space Model (SSM) principles with Transformer layers, employing efficient gated convolutions to achieve linear-time complexity. This hybrid design eliminates the heavy memory I/O overhead of Key-Value caching during generation. Our approach delivers substantial efficiency gains, reducing quantization error of vision encoder by up to 7x and end-to-end latency by 14x compared to conventional baselines. The AutoNeural also delivers 3x decoding speed and 4x longer context window than the baseline. We validate these improvements via a real-world automotive case study on the Qualcomm SA8295P SoC, demonstrating real-time performance for cockpit applications. Our results highlight that rethinking model topology specifically for NPU constraints is a prerequisite for robust multi-modal edge intelligence.

Recent advances in vision-language-action (VLA) models have shown promise in integrating image generation with action prediction to improve generalization and reasoning in robot manipulation. However, existing methods are limited to challenging image-based forecasting, which suffers from redundant information and lacks comprehensive and critical world knowledge, including dynamic, spatial and semantic information. To address these limitations, we propose DreamVLA, a novel VLA framework that integrates comprehensive world knowledge forecasting to enable inverse dynamics modeling, thereby establishing a perception-prediction-action loop for manipulation tasks. Specifically, DreamVLA introduces a dynamic-region-guided world knowledge prediction, integrated with the spatial and semantic cues, which provide compact yet comprehensive representations for action planning. This design aligns with how humans interact with the world by first forming abstract multimodal reasoning chains before acting. To mitigate interference among the dynamic, spatial and semantic information during training, we adopt a block-wise structured attention mechanism that masks their mutual attention, preventing information leakage and keeping each representation clean and disentangled. Moreover, to model the conditional distribution over future actions, we employ a diffusion-based transformer that disentangles action representations from shared latent features. Extensive experiments on both real-world and simulation environments demonstrate that DreamVLA achieves 76.7% success rate on real robot tasks and 4.44 average length on the CALVIN ABC-D benchmarks.

Vision-Language-Action (VLA) models adapt large vision-language backbones to map images and instructions to robot actions. However, prevailing VLA decoders either generate actions autoregressively in a fixed left-to-right order or attach continuous diffusion or flow matching heads outside the backbone, demanding specialized training and iterative sampling that hinder a unified, scalable architecture. We present Discrete Diffusion VLA, a single-transformer policy that models discretized action chunks with discrete diffusion and is trained with the same cross-entropy objective as the VLM backbone. The design retains diffusion's progressive refinement paradigm while remaining natively compatible with the discrete token interface of VLMs. Our method achieves an adaptive decoding order that resolves easy action elements before harder ones and uses secondary remasking to revisit uncertain predictions across refinement rounds, which improves consistency and enables robust error correction. This unified decoder preserves pretrained vision language priors, supports parallel decoding, breaks the autoregressive bottleneck, and reduces the number of function evaluations. Discrete Diffusion VLA achieves 96.3% avg. SR on LIBERO, 71.2% visual matching on SimplerEnv Fractal and 49.3% overall on SimplerEnv Bridge, improving over both autoregressive and continuous diffusion baselines. These findings indicate that discrete-diffusion action decoder supports precise action modeling and consistent training, laying groundwork for scaling VLA to larger models and datasets.

This paper presents PaLI-3, a smaller, faster, and stronger vision language model (VLM) that compares favorably to similar models that are 10x larger. As part of arriving at this strong performance, we compare Vision Transformer (ViT) models pretrained using classification objectives to contrastively (SigLIP) pretrained ones. We find that, while slightly underperforming on standard image classification benchmarks, SigLIP-based PaLI shows superior performance across various multimodal benchmarks, especially on localization and visually-situated text understanding. We scale the SigLIP image encoder up to 2 billion parameters, and achieves a new state-of-the-art on multilingual cross-modal retrieval. We hope that PaLI-3, at only 5B parameters, rekindles research on fundamental pieces of complex VLMs, and could fuel a new generation of scaled-up models.

Autoregressive sequence models, such as Transformer-based vision-language action (VLA) policies, can be tremendously effective for capturing complex and generalizable robotic behaviors. However, such models require us to choose a tokenization of our continuous action signals, which determines how the discrete symbols predicted by the model map to continuous robot actions. We find that current approaches for robot action tokenization, based on simple per-dimension, per-timestep binning schemes, typically perform poorly when learning dexterous skills from high-frequency robot data. To address this challenge, we propose a new compression-based tokenization scheme for robot actions, based on the discrete cosine transform. Our tokenization approach, Frequency-space Action Sequence Tokenization (FAST), enables us to train autoregressive VLAs for highly dexterous and high-frequency tasks where standard discretization methods fail completely. Based on FAST, we release FAST+, a universal robot action tokenizer, trained on 1M real robot action trajectories. It can be used as a black-box tokenizer for a wide range of robot action sequences, with diverse action spaces and control frequencies. Finally, we show that, when combined with the pi0 VLA, our method can scale to training on 10k hours of robot data and match the performance of diffusion VLAs, while reducing training time by up to 5x.

Recent advances in Vision-Language-Action (VLA) models demonstrate that visual signals can effectively complement sparse action supervisions. However, letting VLA directly predict high-dimensional visual states can distribute model capacity and incur prohibitive training cost, while compressing visual states into more compact supervisory signals inevitably incurs information bottlenecks. Moreover, existing methods often suffer from poor comprehension and reasoning capabilities due to the neglect of language supervision. This paper introduces Mantis, a novel framework featuring a Disentangled Visual Foresight (DVF) to tackle these issues. Specifically, Mantis decouples visual foresight prediction from the backbone with the combination of meta queries and a diffusion Transformer (DiT) head. With the current visual state provided to the DiT via a residual connection, a simple next-state prediction objective enables the meta queries to automatically capture the latent actions that delineate the visual trajectory, and hence boost the learning of explicit actions. The disentanglement reduces the burden of the VLA backbone, enabling it to maintain comprehension and reasoning capabilities through language supervision. Empirically, pretrained on human manipulation videos, robot demonstrations, and image-text pairs, Mantis achieves a 96.7% success rate on LIBERO benchmark after fine-tuning, surpassing powerful baselines while exhibiting high convergence speed. Real-world evaluations show that Mantis outperforms π_{0.5}, a leading open-source VLA model, particularly in instruction-following capability, generalization to unseen instructions, and reasoning ability. Code and weights are released to support the open-source community.

Recent advancements in multi-modal models have significantly improved vision-language alignment in radiology. However, existing approaches struggle to effectively utilize complex radiology reports for learning, rely on low-resolution images, and offer limited interpretability in attention mechanisms. To address these challenges, we introduce RadZero, a novel similarity-based cross-attention framework for vision-language alignment in radiology with zero-shot multi-task capability. RadZero leverages large language models to extract minimal semantic sentences from radiology reports and employs a multi-positive contrastive learning strategy to effectively capture relationships between images and multiple relevant textual descriptions. It also utilizes a pre-trained vision encoder with additional trainable Transformer layers, allowing efficient high-resolution image processing. By computing similarity between text embeddings and local image patch features, RadZero enables zero-shot inference with similarity probability for classification and pixel-level cross-modal similarity maps for grounding and segmentation. Experimental results on public chest radiograph benchmarks show that RadZero outperforms state-of-the-art methods in zero-shot classification, grounding, and segmentation. Furthermore, cross-modal similarity map analysis highlights its potential for improving explainability in vision-language alignment. Additionally, qualitative evaluation demonstrates RadZero's capability for open-vocabulary semantic segmentation, further validating its effectiveness in medical imaging.

Large-scale pretraining and instruction tuning have been successful for training general-purpose language models with broad competencies. However, extending to general-purpose vision-language models is challenging due to the distributional diversity in visual inputs. A recent line of work explores vision-language instruction tuning, taking inspiration from the Query Transformer (QFormer) approach proposed in BLIP-2 models for bridging frozen modalities. However, these approaches rely heavily on large-scale multi-modal pretraining for representation learning before eventual finetuning, incurring a huge computational overhead, poor scaling, and limited accessibility. To that end, we propose a more efficient method for QFormer-based vision-language alignment and demonstrate the effectiveness of our strategy compared to existing baselines in improving the efficiency of vision-language pretraining.

Recent advances in Vision-Language-Action (VLA) models have enabled robotic agents to integrate multimodal understanding with action execution. However, our empirical analysis reveals that current VLAs struggle to allocate visual attention to target regions. Instead, visual attention is always dispersed. To guide the visual attention grounding on the correct target, we propose ReconVLA, a reconstructive VLA model with an implicit grounding paradigm. Conditioned on the model's visual outputs, a diffusion transformer aims to reconstruct the gaze region of the image, which corresponds to the target manipulated objects. This process prompts the VLA model to learn fine-grained representations and accurately allocate visual attention, thus effectively leveraging task-specific visual information and conducting precise manipulation. Moreover, we curate a large-scale pretraining dataset comprising over 100k trajectories and 2 million data samples from open-source robotic datasets, further boosting the model's generalization in visual reconstruction. Extensive experiments in simulation and the real world demonstrate the superiority of our implicit grounding method, showcasing its capabilities of precise manipulation and generalization. Our project page is https://zionchow.github.io/ReconVLA/.

Recent advancements enlarge the capabilities of large language models (LLMs) in zero-shot image-to-text generation and understanding by integrating multi-modal inputs. However, such success is typically limited to English scenarios due to the lack of large-scale and high-quality non-English multi-modal resources, making it extremely difficult to establish competitive counterparts in other languages. In this paper, we introduce the Ziya-VL series, a set of bilingual large-scale vision-language models (LVLMs) designed to incorporate visual semantics into LLM for multi-modal dialogue. Composed of Ziya-VL-Base and Ziya-VL-Chat, our models adopt the Querying Transformer from BLIP-2, further exploring the assistance of optimization schemes such as instruction tuning, multi-stage training and low-rank adaptation module for visual-language alignment. In addition, we stimulate the understanding ability of GPT-4 in multi-modal scenarios, translating our gathered English image-text datasets into Chinese and generating instruction-response through the in-context learning method. The experiment results demonstrate that compared to the existing LVLMs, Ziya-VL achieves competitive performance across a wide range of English-only tasks including zero-shot image-text retrieval, image captioning, and visual question answering. The evaluation leaderboard accessed by GPT-4 also indicates that our models possess satisfactory image-text understanding and generation capabilities in Chinese multi-modal scenario dialogues. Code, demo and models are available at ~https://huggingface.co/IDEA-CCNL/Ziya-BLIP2-14B-Visual-v1.

Prompt tuning is a parameter-efficient way to deploy large-scale pre-trained models to downstream tasks by adding task-specific tokens. In terms of vision-language pre-trained (VLP) models, prompt tuning often requires a large number of learnable tokens to bridge the gap between the pre-training and downstream tasks, which greatly exacerbates the already high computational overhead. In this paper, we revisit the principle of prompt tuning for Transformer-based VLP models, and reveal that the impact of soft prompt tokens can be actually approximated via independent information diffusion steps, thereby avoiding the expensive global attention modeling and reducing the computational complexity to a large extent. Based on this finding, we propose a novel Approximated Prompt Tuning (APT) approach towards efficient VL transfer learning. To validate APT, we apply it to two representative VLP models, namely ViLT and METER, and conduct extensive experiments on a bunch of downstream tasks. Meanwhile, the generalization of APT is also validated on CLIP for image classification and StableDiffusion for text-to-image generation. The experimental results not only show the superior performance gains and computation efficiency of APT against the conventional prompt tuning methods, e.g., +7.01% accuracy and -82.30% additional computation overhead on METER, but also confirm its merits over other parameter-efficient transfer learning approaches.

Vision-Language (VL) models have garnered considerable research interest; however, they still face challenges in effectively handling text within images. To address this limitation, researchers have developed two approaches. The first method involves utilizing external Optical Character Recognition (OCR) tools to extract textual information from images, which is then prepended to other textual inputs. The second strategy focuses on employing extremely high-resolution images to improve text recognition capabilities. In this paper, we focus on enhancing the first strategy by introducing a novel method, named TAP-VL, which treats OCR information as a distinct modality and seamlessly integrates it into any VL model. TAP-VL employs a lightweight transformer-based OCR module to receive OCR with layout information, compressing it into a short fixed-length sequence for input into the LLM. Initially, we conduct model-agnostic pretraining of the OCR module on unlabeled documents, followed by its integration into any VL architecture through brief fine-tuning. Extensive experiments demonstrate consistent performance improvements when applying TAP-VL to top-performing VL models, across scene-text and document-based VL benchmarks.

This paper presents a detailed study of improving visual representations for vision language (VL) tasks and develops an improved object detection model to provide object-centric representations of images. Compared to the most widely used bottom-up and top-down model anderson2018bottom, the new model is bigger, better-designed for VL tasks, and pre-trained on much larger training corpora that combine multiple public annotated object detection datasets. Therefore, it can generate representations of a richer collection of visual objects and concepts. While previous VL research focuses mainly on improving the vision-language fusion model and leaves the object detection model improvement untouched, we show that visual features matter significantly in VL models. In our experiments we feed the visual features generated by the new object detection model into a Transformer-based VL fusion model \oscar li2020oscar, and utilize an improved approach \short\ to pre-train the VL model and fine-tune it on a wide range of downstream VL tasks. Our results show that the new visual features significantly improve the performance across all VL tasks, creating new state-of-the-art results on seven public benchmarks. We will release the new object detection model to public.

The growing interest in vision-language models (VLMs) has been driven by improvements in large language models and vision transformers. Despite the abundance of literature on this subject, we observe that critical decisions regarding the design of VLMs are often not justified. We argue that these unsupported decisions impede progress in the field by making it difficult to identify which choices improve model performance. To address this issue, we conduct extensive experiments around pre-trained models, architecture choice, data, and training methods. Our consolidation of findings includes the development of Idefics2, an efficient foundational VLM of 8 billion parameters. Idefics2 achieves state-of-the-art performance within its size category across various multimodal benchmarks, and is often on par with models four times its size. We release the model (base, instructed, and chat) along with the datasets created for its training.

Open-Vocabulary Segmentation (OVS) aims at segmenting images from free-form textual concepts without predefined training classes. While existing vision-language models such as CLIP can generate segmentation masks by leveraging coarse spatial information from Vision Transformers, they face challenges in spatial localization due to their global alignment of image and text features. Conversely, self-supervised visual models like DINO excel in fine-grained visual encoding but lack integration with language. To bridge this gap, we present Talk2DINO, a novel hybrid approach that combines the spatial accuracy of DINOv2 with the language understanding of CLIP. Our approach aligns the textual embeddings of CLIP to the patch-level features of DINOv2 through a learned mapping function without the need to fine-tune the underlying backbones. At training time, we exploit the attention maps of DINOv2 to selectively align local visual patches with textual embeddings. We show that the powerful semantic and localization abilities of Talk2DINO can enhance the segmentation process, resulting in more natural and less noisy segmentations, and that our approach can also effectively distinguish foreground objects from the background. Experimental results demonstrate that Talk2DINO achieves state-of-the-art performance across several unsupervised OVS benchmarks. Source code and models are publicly available at: https://lorebianchi98.github.io/Talk2DINO/.

Compositional reasoning is a hallmark of human visual intelligence; yet despite the size of large vision-language models, they struggle to represent simple compositions by combining objects with their attributes. To measure this lack of compositional capability, we design Cola, a text-to-image retrieval benchmark to Compose Objects Localized with Attributes. Using Cola as a testbed, we explore modeling designs to adapt pre-trained vision-language models to reason compositionally about multiple attributes attached to multiple objects. We explore 6 finetuning strategies on 2 seminal vision-language models, using 3 finetuning datasets and 2 test benchmarks (Cola and CREPE). Surprisingly, our optimal finetuning strategy improves a 151M parameter CLIP, which disjointly encodes image and language during pretraining, to perform as well as a 241M parameter FLAVA, which uses a multi-modal transformer encoder during pretraining to attend over both vision and language modalities. This optimal finetuning strategy is a lightweight multi-modal adapter that jointly attends over both image and language features generated by the pretrained model. We show this works better than common strategies such as prompt/fine-tuning, or tuning a comparable number of unimodal layers.

Vision-Language Models (VLMs) have emerged as powerful tools in artificial intelli-gence, capable of integrating textual and visual data for a unified understanding of complex scenes. While models such as Florence2, built on transformer architectures, have shown promise across general tasks, their performance in object detection within unstructured or cluttered environments remains underexplored. In this study, we fi-ne-tuned the Florence2 model for object detection tasks in non-constructed, complex environments. A comprehensive experimental framework was established involving multiple hardware configurations (NVIDIA T4, L4, and A100 GPUs), optimizers (AdamW, SGD), and varied hyperparameters including learning rates and LoRA (Low-Rank Adaptation) setups. Model training and evaluation were conducted on challenging datasets representative of real-world, disordered settings. The optimized Florence2 models exhibited significant improvements in object detection accuracy, with Mean Average Precision (mAP) metrics approaching or matching those of estab-lished models such as YOLOv8, YOLOv9, and YOLOv10. The integration of LoRA and careful fine-tuning of transformer layers contributed notably to these gains. Our find-ings highlight the adaptability of transformer-based VLMs like Florence2 for do-main-specific tasks, particularly in visually complex environments. The study under-scores the potential of fine-tuned VLMs to rival traditional convolution-based detec-tors, offering a flexible and scalable approach for advanced vision applications in re-al-world, unstructured settings.

In the past several years there has been an explosion of available models for vision-language tasks. Unfortunately, the literature still leaves open a number of questions related to best practices in designing and training such models. In this paper we seek to answer several questions related to the pretraining of vision-language encoders through meta-analysis. In our first set of experiments, we show that we can save significant compute at no cost to downstream performance, by freezing large parts of vision-language models during pretraining. In our second set of experiments we examine the effect of basing a VL transformer on a vision model versus a text model. Additionally, we introduce a VL modeling platform called Renaissance that we use to conduct all of the experiments. This program offers a great deal of flexibility in creating, training and evaluating transformer encoders for VL modeling. The source code for Renaissance can be found at https://github.com/bsu-slim/renaissance.

Foundation Vision-Language Models (VLMs) trained using large-scale open-domain images and text pairs have recently been adapted to develop Vision-Language Segmentation Models (VLSMs) that allow providing text prompts during inference to guide image segmentation. If robust and powerful VLSMs can be built for medical images, it could aid medical professionals in many clinical tasks where they must spend substantial time delineating the target structure of interest. VLSMs for medical images resort to fine-tuning base VLM or VLSM pretrained on open-domain natural image datasets due to fewer annotated medical image datasets; this fine-tuning is resource-consuming and expensive as it usually requires updating all or a significant fraction of the pretrained parameters. Recently, lightweight blocks called adapters have been proposed in VLMs that keep the pretrained model frozen and only train adapters during fine-tuning, substantially reducing the computing resources required. We introduce a novel adapter, VLSM-Adapter, that can fine-tune pretrained vision-language segmentation models using transformer encoders. Our experiments in widely used CLIP-based segmentation models show that with only 3 million trainable parameters, the VLSM-Adapter outperforms state-of-the-art and is comparable to the upper bound end-to-end fine-tuning. The source code is available at: https://github.com/naamiinepal/vlsm-adapter.

Background: General-purpose large language models that utilize both text and images have not been evaluated on a diverse array of challenging medical cases. Methods: Using 934 cases from the NEJM Image Challenge published between 2005 and 2023, we evaluated the accuracy of the recently released Generative Pre-trained Transformer 4 with Vision model (GPT-4V) compared to human respondents overall and stratified by question difficulty, image type, and skin tone. We further conducted a physician evaluation of GPT-4V on 69 NEJM clinicopathological conferences (CPCs). Analyses were conducted for models utilizing text alone, images alone, and both text and images. Results: GPT-4V achieved an overall accuracy of 61% (95% CI, 58 to 64%) compared to 49% (95% CI, 49 to 50%) for humans. GPT-4V outperformed humans at all levels of difficulty and disagreement, skin tones, and image types; the exception was radiographic images, where performance was equivalent between GPT-4V and human respondents. Longer, more informative captions were associated with improved performance for GPT-4V but similar performance for human respondents. GPT-4V included the correct diagnosis in its differential for 80% (95% CI, 68 to 88%) of CPCs when using text alone, compared to 58% (95% CI, 45 to 70%) of CPCs when using both images and text. Conclusions: GPT-4V outperformed human respondents on challenging medical cases and was able to synthesize information from both images and text, but performance deteriorated when images were added to highly informative text. Overall, our results suggest that multimodal AI models may be useful in medical diagnostic reasoning but that their accuracy may depend heavily on context.

Vision-Language-Action (VLA) models aim to predict robotic actions based on visual observations and language instructions. Existing approaches require fine-tuning pre-trained visionlanguage models (VLMs) as visual and language features are independently fed into downstream policies, degrading the pre-trained semantic alignments. We propose OTTER, a novel VLA architecture that leverages these existing alignments through explicit, text-aware visual feature extraction. Instead of processing all visual features, OTTER selectively extracts and passes only task-relevant visual features that are semantically aligned with the language instruction to the policy transformer. This allows OTTER to keep the pre-trained vision-language encoders frozen. Thereby, OTTER preserves and utilizes the rich semantic understanding learned from large-scale pre-training, enabling strong zero-shot generalization capabilities. In simulation and real-world experiments, OTTER significantly outperforms existing VLA models, demonstrating strong zeroshot generalization to novel objects and environments. Video, code, checkpoints, and dataset: https://ottervla.github.io/.

The Large Vision-Language Model (LVLM) field has seen significant advancements, yet its progression has been hindered by challenges in comprehending fine-grained visual content due to limited resolution. Recent efforts have aimed to enhance the high-resolution understanding capabilities of LVLMs, yet they remain capped at approximately 1500 x 1500 pixels and constrained to a relatively narrow resolution range. This paper represents InternLM-XComposer2-4KHD, a groundbreaking exploration into elevating LVLM resolution capabilities up to 4K HD (3840 x 1600) and beyond. Concurrently, considering the ultra-high resolution may not be necessary in all scenarios, it supports a wide range of diverse resolutions from 336 pixels to 4K standard, significantly broadening its scope of applicability. Specifically, this research advances the patch division paradigm by introducing a novel extension: dynamic resolution with automatic patch configuration. It maintains the training image aspect ratios while automatically varying patch counts and configuring layouts based on a pre-trained Vision Transformer (ViT) (336 x 336), leading to dynamic training resolution from 336 pixels to 4K standard. Our research demonstrates that scaling training resolution up to 4K HD leads to consistent performance enhancements without hitting the ceiling of potential improvements. InternLM-XComposer2-4KHD shows superb capability that matches or even surpasses GPT-4V and Gemini Pro in 10 of the 16 benchmarks. The InternLM-XComposer2-4KHD model series with 7B parameters are publicly available at https://github.com/InternLM/InternLM-XComposer.

Modern Large Vision-Language Models (LVLMs) enjoy the same vision vocabulary -- CLIP, which can cover most common vision tasks. However, for some special vision task that needs dense and fine-grained vision perception, e.g., document-level OCR or chart understanding, especially in non-English scenarios, the CLIP-style vocabulary may encounter low efficiency in tokenizing the vision knowledge and even suffer out-of-vocabulary problem. Accordingly, we propose Vary, an efficient and effective method to scale up the vision vocabulary of LVLMs. The procedures of Vary are naturally divided into two folds: the generation and integration of a new vision vocabulary. In the first phase, we devise a vocabulary network along with a tiny decoder-only transformer to produce the desired vocabulary via autoregression. In the next, we scale up the vanilla vision vocabulary by merging the new one with the original one (CLIP), enabling the LVLMs can quickly garner new features. Compared to the popular BLIP-2, MiniGPT4, and LLaVA, Vary can maintain its vanilla capabilities while enjoying more excellent fine-grained perception and understanding ability. Specifically, Vary is competent in new document parsing features (OCR or markdown conversion) while achieving 78.2% ANLS in DocVQA and 36.2% in MMVet. Our code will be publicly available on the homepage.

Large Vision Language Models (LVLMs) have demonstrated remarkable capabilities in understanding and describing visual content, achieving state-of-the-art performance across various vision-language tasks. However, these models frequently exhibit hallucination behavior, where they generate descriptions containing objects or details absent in the input image. Our work investigates this phenomenon by analyzing attention patterns across transformer layers and heads, revealing that hallucinations often stem from progressive degradation of visual grounding in deeper layers. We propose a novel attention modification approach that combines selective token emphasis and head-specific modulation to maintain visual grounding throughout the generation process. Our method introduces two key components: (1) a dual-stream token selection mechanism that identifies and prioritizes both locally informative and spatially significant visual tokens, and (2) an attention head-specific modulation strategy that differentially amplifies visual information processing based on measured visual sensitivity of individual attention heads. Through extensive experimentation on the MSCOCO dataset, we demonstrate that our approach reduces hallucination rates by up to 62.3\% compared to baseline models while maintaining comparable task performance. Our analysis reveals that selectively modulating tokens across attention heads with varying levels of visual sensitivity can significantly improve visual grounding without requiring model retraining.

Vision Transformer has shown great visual representation power in substantial vision tasks such as recognition and detection, and thus been attracting fast-growing efforts on manually designing more effective architectures. In this paper, we propose to use neural architecture search to automate this process, by searching not only the architecture but also the search space. The central idea is to gradually evolve different search dimensions guided by their E-T Error computed using a weight-sharing supernet. Moreover, we provide design guidelines of general vision transformers with extensive analysis according to the space searching process, which could promote the understanding of vision transformer. Remarkably, the searched models, named S3 (short for Searching the Search Space), from the searched space achieve superior performance to recently proposed models, such as Swin, DeiT and ViT, when evaluated on ImageNet. The effectiveness of S3 is also illustrated on object detection, semantic segmentation and visual question answering, demonstrating its generality to downstream vision and vision-language tasks. Code and models will be available at https://github.com/microsoft/Cream.

Different medical imaging modalities capture diagnostic information at varying spatial resolutions, from coarse global patterns to fine-grained localized structures. However, most existing vision-language frameworks in the medical domain apply a uniform strategy for local feature extraction, overlooking the modality-specific demands. In this work, we present MedMoE, a modular and extensible vision-language processing framework that dynamically adapts visual representation based on the diagnostic context. MedMoE incorporates a Mixture-of-Experts (MoE) module conditioned on the report type, which routes multi-scale image features through specialized expert branches trained to capture modality-specific visual semantics. These experts operate over feature pyramids derived from a Swin Transformer backbone, enabling spatially adaptive attention to clinically relevant regions. This framework produces localized visual representations aligned with textual descriptions, without requiring modality-specific supervision at inference. Empirical results on diverse medical benchmarks demonstrate that MedMoE improves alignment and retrieval performance across imaging modalities, underscoring the value of modality-specialized visual representations in clinical vision-language systems.

Robot vision has greatly benefited from advancements in multimodal fusion techniques and vision-language models (VLMs). We systematically review the applications of multimodal fusion in key robotic vision tasks, including semantic scene understanding, simultaneous localization and mapping (SLAM), 3D object detection, navigation and localization, and robot manipulation. We compare VLMs based on large language models (LLMs) with traditional multimodal fusion methods, analyzing their advantages, limitations, and synergies. Additionally, we conduct an in-depth analysis of commonly used datasets, evaluating their applicability and challenges in real-world robotic scenarios. Furthermore, we identify critical research challenges such as cross-modal alignment, efficient fusion strategies, real-time deployment, and domain adaptation, and propose future research directions, including self-supervised learning for robust multimodal representations, transformer-based fusion architectures, and scalable multimodal frameworks. Through a comprehensive review, comparative analysis, and forward-looking discussion, we provide a valuable reference for advancing multimodal perception and interaction in robotic vision. A comprehensive list of studies in this survey is available at https://github.com/Xiaofeng-Han-Res/MF-RV.

Leveraging sensing modalities across diverse spatial and temporal resolutions can improve performance of robotic manipulation tasks. Multi-spatial resolution sensing provides hierarchical information captured at different spatial scales and enables both coarse and precise motions. Simultaneously multi-temporal resolution sensing enables the agent to exhibit high reactivity and real-time control. In this work, we propose a framework, MResT (Multi-Resolution Transformer), for learning generalizable language-conditioned multi-task policies that utilize sensing at different spatial and temporal resolutions using networks of varying capacities to effectively perform real time control of precise and reactive tasks. We leverage off-the-shelf pretrained vision-language models to operate on low-frequency global features along with small non-pretrained models to adapt to high frequency local feedback. Through extensive experiments in 3 domains (coarse, precise and dynamic manipulation tasks), we show that our approach significantly improves (2X on average) over recent multi-task baselines. Further, our approach generalizes well to visual and geometric variations in target objects and to varying interaction forces.

Window attention and linear attention represent two principal strategies for mitigating the quadratic complexity and ever-growing KV cache in Vision-Language Models (VLMs). However, we observe that window-based VLMs suffer performance degradation when sequence length exceeds the window size, while linear attention underperforms on information-intensive tasks such as OCR and document understanding. To overcome these limitations, we propose InfiniteVL, a linear-complexity VLM architecture that synergizes sliding window attention (SWA) with Gated DeltaNet. For achieving competitive multimodal performance under constrained resources, we design a three-stage training strategy comprising distillation pretraining, instruction tuning, and long-sequence SFT. Remarkably, using less than 2\% of the training data required by leading VLMs, InfiniteVL not only substantially outperforms previous linear-complexity VLMs but also matches the performance of leading Transformer-based VLMs, while demonstrating effective long-term memory retention. Compared to similar-sized Transformer-based VLMs accelerated by FlashAttention-2, InfiniteVL achieves over 3.6\times inference speedup while maintaining constant latency and memory footprint. In streaming video understanding scenarios, it sustains a stable 24 FPS real-time prefill speed while preserving long-term memory cache. Code and models are available at https://github.com/hustvl/InfiniteVL.

Despite significant advancements in Multimodal Large Language Models (MLLMs) for understanding complex human intentions through cross-modal interactions, capturing intricate image details remains challenging. Previous methods integrating multiple vision encoders to enhance visual detail introduce redundancy and computational overhead. We observe that most MLLMs utilize only the last-layer feature map of the vision encoder for visual representation, neglecting the rich fine-grained information in shallow feature maps. To address this issue, we propose \modelname, a simple yet effective multi-layer feature fuser that efficiently integrates deep and shallow features from Vision Transformers (ViTs). Specifically, it leverages semantically aligned deep features as queries to dynamically extract missing details from shallow features, thus preserving semantic alignment while enriching the representation with fine-grained information. Applied to the LLaVA-1.5 model, \modelname~achieves significant improvements in visual representation and benchmark performance, providing a more flexible and lightweight solution compared to multi-encoder ensemble methods. The code and model have been released at https://github.com/yuecao0119/MMFuser.

We present Unified-IO 2, the first autoregressive multimodal model that is capable of understanding and generating image, text, audio, and action. To unify different modalities, we tokenize inputs and outputs -- images, text, audio, action, bounding boxes, etc., into a shared semantic space and then process them with a single encoder-decoder transformer model. Since training with such diverse modalities is challenging, we propose various architectural improvements to stabilize model training. We train our model from scratch on a large multimodal pre-training corpus from diverse sources with a multimodal mixture of denoisers objective. To learn an expansive set of skills, such as following multimodal instructions, we construct and finetune on an ensemble of 120 datasets with prompts and augmentations. With a single unified model, Unified-IO 2 achieves state-of-the-art performance on the GRIT benchmark and strong results in more than 35 benchmarks, including image generation and understanding, natural language understanding, video and audio understanding, and robotic manipulation. We release all our models to the research community.

Large Vision-Language-Action (VLA) models, leveraging powerful pre trained Vision-Language Models (VLMs) backends, have shown promise in robotic control due to their impressive generalization ability. However, the success comes at a cost. Their reliance on VLM backends with billions of parameters leads to high computational costs and inference latency, limiting the testing scenarios to mainly quasi-static tasks and hindering performance in dynamic tasks requiring rapid interactions. To address these limitations, this paper proposes HiRT, a Hierarchical Robot Transformer framework that enables flexible frequency and performance trade-off. HiRT keeps VLMs running at low frequencies to capture temporarily invariant features while enabling real-time interaction through a high-frequency vision-based policy guided by the slowly updated features. Experiment results in both simulation and real-world settings demonstrate significant improvements over baseline methods. Empirically, in static tasks, we double the control frequency and achieve comparable success rates. Additionally, on novel real-world dynamic ma nipulation tasks which are challenging for previous VLA models, HiRT improves the success rate from 48% to 75%.

Video Highlight Detection and Moment Retrieval (HD/MR) are essential in video analysis. Recent joint prediction transformer models often overlook their cross-task dynamics and video-text alignment and refinement. Moreover, most models typically use limited, uni-directional attention mechanisms, resulting in weakly integrated representations and suboptimal performance in capturing the interdependence between video and text modalities. Although large-language and vision-language models (LLM/LVLMs) have gained prominence across various domains, their application in this field remains relatively underexplored. Here we propose VideoLights, a novel HD/MR framework addressing these limitations through (i) Convolutional Projection and Feature Refinement modules with an alignment loss for better video-text feature alignment, (ii) Bi-Directional Cross-Modal Fusion network for strongly coupled query-aware clip representations, and (iii) Uni-directional joint-task feedback mechanism enhancing both tasks through correlation. In addition, (iv) we introduce hard positive/negative losses for adaptive error penalization and improved learning, and (v) leverage LVLMs like BLIP-2 for enhanced multimodal feature integration and intelligent pretraining using synthetic data generated from LVLMs. Comprehensive experiments on QVHighlights, TVSum, and Charades-STA benchmarks demonstrate state-of-the-art performance. Codes and models are available at https://github.com/dpaul06/VideoLights .

Language and vision-language models have shown impressive performance across a wide range of tasks, but their internal mechanisms remain only partly understood. In this work, we study how individual attention heads in text-generative models specialize in specific semantic or visual attributes. Building on an established interpretability method, we reinterpret the practice of probing intermediate activations with the final decoding layer through the lens of signal processing. This lets us analyze multiple samples in a principled way and rank attention heads based on their relevance to target concepts. Our results show consistent patterns of specialization at the head level across both unimodal and multimodal transformers. Remarkably, we find that editing as few as 1% of the heads, selected using our method, can reliably suppress or enhance targeted concepts in the model output. We validate our approach on language tasks such as question answering and toxicity mitigation, as well as vision-language tasks including image classification and captioning. Our findings highlight an interpretable and controllable structure within attention layers, offering simple tools for understanding and editing large-scale generative models.

In this paper, we study the problem of generalizable synthetic image detection, aiming to detect forgery images from diverse generative methods, e.g., GANs and diffusion models. Cutting-edge solutions start to explore the benefits of pre-trained models, and mainly follow the fixed paradigm of solely training an attached classifier, e.g., combining frozen CLIP-ViT with a learnable linear layer in UniFD. However, our analysis shows that such a fixed paradigm is prone to yield detectors with insufficient learning regarding forgery representations. We attribute the key challenge to the lack of forgery adaptation, and present a novel forgery-aware adaptive transformer approach, namely FatFormer. Based on the pre-trained vision-language spaces of CLIP, FatFormer introduces two core designs for the adaption to build generalized forgery representations. First, motivated by the fact that both image and frequency analysis are essential for synthetic image detection, we develop a forgery-aware adapter to adapt image features to discern and integrate local forgery traces within image and frequency domains. Second, we find that considering the contrastive objectives between adapted image features and text prompt embeddings, a previously overlooked aspect, results in a nontrivial generalization improvement. Accordingly, we introduce language-guided alignment to supervise the forgery adaptation with image and text prompts in FatFormer. Experiments show that, by coupling these two designs, our approach tuned on 4-class ProGAN data attains a remarkable detection performance, achieving an average of 98% accuracy to unseen GANs, and surprisingly generalizes to unseen diffusion models with 95% accuracy.