The cold-start initialization stage plays a pivotal role in training Multimodal Large Reasoning Models (MLRMs), yet its mechanisms remain insufficiently understood. To analyze this stage, we introduce the Visual Attention Score (VAS), an attention-based metric that quantifies how much a model attends to visual tokens. We find that reasoning performance is strongly correlated with VAS (r=0.9616): models with higher VAS achieve substantially stronger multimodal reasoning. Surprisingly, multimodal cold-start fails to raise VAS, leaving distributions close to the base model, whereas text-only cold-start induces a clear increase. We term this counter-intuitive phenomenon Lazy Attention Localization. To validate its causal role, we design training-free interventions that directly manipulate attention allocation at inference time, yielding consistent 1--2% gains without retraining. Building on these insights, we propose Attention-Guided Visual Anchoring and Reflection (AVAR), a comprehensive cold-start framework that integrates visual-anchored data synthesis, attention-guided objectives, and visual-anchored reward shaping. Applied to Qwen2.5-VL-7B, AVAR delivers an average gain of 7.0% across 7 multimodal reasoning benchmarks. Ablation studies further confirm that each component of AVAR contributes step-wise to the overall gains.

We present Qwen-Drive-1.0, an initial step towards a vision-language foundation model for autonomous driving. Qwen-Drive-1.0 retains the architecture of the pretrained vision-language model (VLM) and integrates 3D perception, visual question answering, and motion planning within a unified framework. An external bird's-eye-view (BEV) perception head jointly performs 3D object detection, semantic occupancy prediction, and BEV map segmentation. It serves as a probe of the 3D information accessible from the shared representations and provides an explicit, inspectable interface to 3D scene structure. A Planning Expert conditions on shared VLM representations to generate future ego trajectories. A staged training recipe combines driving supervision with general-purpose vision-language data to acquire driving-specific competence while helping preserve broad visual understanding and instruction-following capabilities. Experiments demonstrate strong 3D perception and driving scene understanding while largely preserving general vision-language capability. Comprehensive evaluations across open-loop, pseudo-closed-loop, and closed-loop settings further show highly competitive motion-planning performance.

When MLLMs fail at Science, Technology, Engineering, and Mathematics (STEM) visual reasoning, a fundamental question arises: is it due to perceptual deficiencies or reasoning limitations? Through systematic scaling analysis that independently scales perception and reasoning components, we uncover a critical insight: scaling perception consistently outperforms scaling reasoning. This reveals perception as the true lever limiting current STEM visual reasoning. Motivated by this insight, our work focuses on systematically enhancing the perception capabilities of MLLMs by establishing code as a powerful perceptual medium—executable code provides precise semantics that naturally align with the structured nature of STEM visuals. Specifically, we construct ICC-1M, a large-scale dataset comprising 1M Image-Caption-Code triplets that materializes this code-as-perception paradigm through two complementary approaches: (1) Code-Grounded Caption Generation treats executable code as ground truth for image captions, eliminating the hallucinations inherent in existing knowledge distillation methods; (2) STEM Image-to-Code Translation prompts models to generate reconstruction code, mitigating the ambiguity of natural language for perception enhancement. To validate this paradigm, we further introduce STEM2Code-Eval, a novel benchmark that directly evaluates visual perception in STEM domains. Unlike existing work relying on problem-solving accuracy as a proxy that only measures problem-relevant understanding, our benchmark requires comprehensive visual comprehension through executable code generation for image reconstruction, providing deterministic and verifiable assessment. Code will be available soon.

We introduce Qwen2.5-VL, the latest flagship model of Qwen vision-language series, which demonstrates significant advancements in both foundational capabilities and innovative functionalities. Qwen2.5-VL achieves a major leap forward in understanding and interacting with the world through enhanced visual recognition, precise object localization, robust document parsing, and long-video comprehension. A standout feature of Qwen2.5-VL is its ability to localize objects using bounding boxes or points accurately. It provides robust structured data extraction from invoices, forms, and tables, as well as detailed analysis of charts, diagrams, and layouts. To handle complex inputs, Qwen2.5-VL introduces dynamic resolution processing and absolute time encoding, enabling it to process images of varying sizes and videos of extended durations (up to hours) with second-level event localization. This allows the model to natively perceive spatial scales and temporal dynamics without relying on traditional normalization techniques. By training a native dynamic-resolution Vision Transformer (ViT) from scratch and incorporating Window Attention, we reduce computational overhead while maintaining native resolution. As a result, Qwen2.5-VL excels not only in static image and document understanding but also as an interactive visual agent capable of reasoning, tool usage, and task execution in real-world scenarios such as operating computers and mobile devices. Qwen2.5-VL is available in three sizes, addressing diverse use cases from edge AI to high-performance computing. The flagship Qwen2.5-VL-72B model matches state-of-the-art models like GPT-4o and Claude 3.5 Sonnet, particularly excelling in document and diagram understanding. Additionally, Qwen2.5-VL maintains robust linguistic performance, preserving the core language competencies of the Qwen2.5 LLM.

We introduce Qwen3-VL, the most capable vision-language model in the Qwen series to date, achieving superior performance across a broad range of multimodal benchmarks. It natively supports interleaved contexts of up to 256K tokens, seamlessly integrating text, images, and video. The model family includes both dense (2B/4B/8B/32B) and mixture-of-experts (30B-A3B/235B-A22B) variants to accommodate diverse latency-quality trade-offs. Qwen3-VL delivers three core pillars: (i) markedly stronger pure-text understanding, surpassing comparable text-only backbones in several cases; (ii) robust long-context comprehension with a native 256K-token window for both text and interleaved multimodal inputs, enabling faithful retention, retrieval, and cross-referencing across long documents and videos; and (iii) advanced multimodal reasoning across single-image, multi-image, and video tasks, demonstrating leading performance on comprehensive evaluations such as MMMU and visual-math benchmarks (e.g., MathVista and MathVision). Architecturally, we introduce three key upgrades: (i) an enhanced interleaved-MRoPE for stronger spatial-temporal modeling across images and video; (ii) DeepStack integration, which effectively leverages multi-level ViT features to tighten vision-language alignment; and (iii) text-based time alignment for video, evolving from T-RoPE to explicit textual timestamp alignment for more precise temporal grounding. Under comparable token budgets and latency constraints, Qwen3-VL achieves superior performance in both dense and Mixture-of-Experts (MoE) architectures. We envision Qwen3-VL serving as a foundational engine for image-grounded reasoning, agentic decision-making, and multimodal code intelligence in real-world workflows.

We present Qwen3-Omni, a single multimodal model that, for the first time, maintains state-of-the-art performance across text, image, audio, and video without any degradation relative to single-modal counterparts. Qwen3-Omni matches the performance of same-sized single-modal models within the Qwen series and excels particularly on audio tasks. Across 36 audio and audio-visual benchmarks, Qwen3-Omni achieves open-source SOTA on 32 benchmarks and overall SOTA on 22, outperforming strong closed-source models such as Gemini-2.5-Pro, Seed-ASR, and GPT-4o-Transcribe. Qwen3-Omni adopts a Thinker-Talker MoE architecture that unifies perception and generation across text, images, audio, and video, yielding fluent text and natural real-time speech. It supports text interaction in 119 languages, speech understanding in 19 languages, and speech generation in 10 languages. To reduce first-packet latency in streaming synthesis, Talker autoregressively predicts discrete speech codecs using a multi-codebook scheme. Leveraging the representational capacity of these codebooks, we replace computationally intensive block-wise diffusion with a lightweight causal ConvNet, enabling streaming from the first codec frame. In cold-start settings, Qwen3-Omni achieves a theoretical end-to-end first-packet latency of 234 ms. To further strengthen multimodal reasoning, we introduce a Thinking model that explicitly reasons over inputs from any modality. Since the research community currently lacks a general-purpose audio captioning model, we fine-tuned Qwen3-Omni-30B-A3B to obtain Qwen3-Omni-30B-A3B-Captioner, which produces detailed, low-hallucination captions for arbitrary audio inputs. Qwen3-Omni-30B-A3B, Qwen3-Omni-30B-A3B-Thinking, and Qwen3-Omni-30B-A3B-Captioner are publicly released under the Apache 2.0 license.

Large Multimodal Models (LMMs) have demonstrated impressive performance in recognizing document images with natural language instructions. However, it remains unclear to what extent capabilities in literacy with rich structure and fine-grained visual challenges. The current landscape lacks a comprehensive benchmark to effectively measure the literate capabilities of LMMs. Existing benchmarks are often limited by narrow scenarios and specified tasks. To this end, we introduce CC-OCR, a comprehensive benchmark that possesses a diverse range of scenarios, tasks, and challenges. CC-OCR comprises four OCR-centric tracks: multi-scene text reading, multilingual text reading, document parsing, and key information extraction. It includes 39 subsets with 7,058 full annotated images, of which 41 for the first time. We evaluate nine prominent LMMs and reveal both the strengths and weaknesses of these models, particularly in text grounding, multi-orientation, and hallucination of repetition. CC-OCR aims to comprehensively evaluate the capabilities of LMMs on OCR-centered tasks, facilitating continued progress in this crucial area.

In recent years, scene text detection has witnessed considerable advancements. However, such methods do not dynamically mine diverse text characteristics within each image to adaptively adjust model parameters, resulting in suboptimal detection performance. To address this issue, we propose a simple yet effective segmentation-based model named Text Query Detector (TQD), inspired by the recently popular transformer. TQD implicitly queries textual information and flexibly generates convolution parameters with the global receptive field. In addition, we decouple the features for parameter generation and dynamic convolution to maximize the benefits of both transformer and convolution. Extensive experiments demonstrate that our approach strikes an ideal tradeoff in terms of both accuracy and speed on prevalent benchmarks. Especially on MSRA-TD500 and ICDAR2015, our TQD achieves state-of-the-art results while maintaining high speed. Code is available at: https://github.com/TangLinJie/TQD.

Scene text recognition (STR) is a challenging task that requires large-scale annotated data for training. However, collecting and labeling real text images is expensive and time-consuming, which limits the availability of real data. Therefore, most existing STR methods resort to synthetic data, which may introduce domain discrepancy and degrade the performance of STR models. To alleviate this problem, recent semi-supervised STR methods exploit unlabeled real data by enforcing character-level consistency regularization between weakly and strongly augmented views of the same image. However, these methods neglect word-level consistency, which is crucial for sequence recognition tasks. This paper proposes a novel semi-supervised learning method for STR that incorporates word-level consistency regularization from both visual and semantic aspects. Specifically, we devise a shortest path alignment module to align the sequential visual features of different views and minimize their distance. Moreover, we adopt a reinforcement learning framework to optimize the semantic similarity of the predicted strings in the embedding space. We conduct extensive experiments on several standard and challenging STR benchmarks and demonstrate the superiority of our proposed method over existing semi-supervised STR methods.

Scene text recognition is a rapidly developing field that faces numerous challenges due to the complexity and diversity of scene text, including complex backgrounds, diverse fonts, flexible arrangements, and accidental occlusions. In this paper, we propose a novel approach called Class-Aware Mask-guided feature refinement (CAM) to address these challenges. Our approach introduces canonical class-aware glyph masks generated from a standard font to effectively suppress background and text style noise, thereby enhancing feature discrimination. Additionally, we design a feature alignment and fusion module to incorporate the canonical mask guidance for further feature refinement for text recognition. By enhancing the alignment between the canonical mask feature and the text feature, the module ensures more effective fusion, ultimately leading to improved recognition performance. We first evaluate CAM on six standard text recognition benchmarks to demonstrate its effectiveness. Furthermore, CAM exhibits superiority over the state-of-the-art method by an average performance gain of 4.1% across six more challenging datasets, despite utilizing a smaller model size. Our study highlights the importance of incorporating canonical mask guidance and aligned feature refinement techniques for robust scene text recognition. Code will be available at https://github.com/MelosY/CAM.

Referring expression comprehension (REC) aims at locating a specific object within a scene given a natural language expression. Although referring expression comprehension has achieved tremendous progress, most of today’s REC models ignore the scene texts in images. Scene text is ubiquitous in our society, and frequently critical to understand the visual scene. To study how to comprehend scene text in the referring expression comprehension task, we collect a novel dataset, termed TextREC, in which most of the referring expressions are related to scene text. Our TextREC dataset challenges a model to recognize scene text, relate it to the referring expressions, and select the most relevant visual object. We also propose a text-guided adaptive modular network (TAMN) to comprehend scene text associated with objects in images. Experimental results reveal that current state-of-the-art REC methods fall short on the TextREC dataset, while our TAMN gets inspiring results by integrating scene text.

Visual information extraction (VIE), which aims to simultaneously perform OCR and information extraction in a unified framework, has drawn increasing attention due to its essential role in various applications like understanding receipts, goods, and traffic signs. However, as existing benchmark datasets for VIE mainly consist of document images without the adequate diversity of layout structures, background disturbs, and entity categories, they cannot fully reveal the challenges of real-world applications. In this paper, we propose a large-scale dataset consisting of camera images for VIE, which contains not only the larger variance of layout, backgrounds, and fonts but also much more types of entities. Besides, we propose a novel framework for end-to-end VIE that combines the stages of OCR and information extraction in an end-to-end learning fashion. Different from the previous end-to-end approaches that directly adopt OCR features as the input of an information extraction module, we propose to use contrastive learning to narrow the semantic gap caused by the difference between the tasks of OCR and information extraction. We evaluate the existing end-to-end methods for VIE on the proposed dataset and observe that the performance of these methods has a distinguishable drop from SROIE (a widely used English dataset) to our proposed dataset due to the larger variance of layout and entities. These results demonstrate our dataset is more practical for promoting advanced VIE algorithms. In addition, experiments demonstrate that the proposed VIE method consistently achieves the obvious performance gains on the proposed and SROIE datasets.

Existing text recognition methods usually need large-scale training data. Most of them rely on synthetic training data due to the lack of annotated real images. However, there is a domain gap between the synthetic data and real data, which limits the performance of the text recognition models. Recent self-supervised text recognition methods attempted to utilize unlabeled real images by introducing contrastive learning, which mainly learns the discrimination of the text images. Inspired by the observation that humans learn to recognize the texts through both reading and writing, we propose to learn discrimination and generation by integrating contrastive learning and masked image modeling in our self-supervised method. The contrastive learning branch is adopted to learn the discrimination of text images, which imitates the reading behavior of humans. Meanwhile, masked image modeling is firstly introduced for text recognition to learn the context generation of the text images, which is similar to the writing behavior. The experimental results show that our method outperforms previous self-supervised text recognition methods by 10.2%-20.2% on irregular scene text recognition datasets. Moreover, our proposed text recognizer exceeds previous state-of-the-art text recognition methods by averagely 5.3% on 11benchmarks, with similar model size. We also demonstrate that our pre-trained model can be easily applied to other text-related tasks with obvious performance gain.

Recently, transformer-based methods have achieved promising progresses in object detection, as they can eliminate the post-processes like NMS and enrich the deep representations. However, these methods cannot well cope with scene text due to its extreme variance of scales and aspect ratios. In this paper, we present a simple yet effective transformer-based architecture for scene text detection. Different from previous approaches that learn robust deep representations of scene text in a holistic manner, our method performs scene text detection based on a few representative features, which avoids the disturbance by background and reduces the computational cost. Specifically, we first select a few representative features at all scales that are highly relevant to foreground text. Then, we adopt a transformer for modeling the relationship of the sampled features, which effectively divides them into reasonable groups. As each feature group corresponds to a text instance, its bounding box can be easily obtained without any post-processing operation. Using the basic feature pyramid network for feature extraction, our method consistently achieves state-of-the-art results on several popular datasets for scene text detection.

OCR-based image captioning aims to automatically describe images based on all the visual entities (both visual objects and scene text) in images. Compared with conventional image captioning, the reasoning of scene text is required for OCR-based image captioning since the generated descriptions often contain multiple OCR tokens. Existing methods attempt to achieve this goal via encoding the OCR tokens with rich visual and semantic representations. However, strong correlations between OCR tokens may not be established with such limited representations. In this paper, we propose to enhance the connections between OCR tokens from the viewpoint of exploiting the geometrical relationship. We comprehensively consider the height, width, distance, IoU and orientation relations between the OCR tokens for constructing the geometrical relationship. To integrate the learned relation as well as the visual and semantic representations into a unified framework, a Long Short-Term Memory plus Relation-aware pointer network (LSTM-R) architecture is presented in this paper. Under the guidance of the geometrical relationship between OCR tokens, our LSTM-R capitalizes on a newly-devised relation-aware pointer network to select OCR tokens from the scene text for OCR-based image captioning. Extensive experiments demonstrate the effectiveness of our LSTM-R. More remarkably, LSTM-R achieves state-of-the-art performance on TextCaps, with the CIDEr-D score being increased from 98.0% to 109.3%.

Scene text retrieval aims to localize and search all text instances from an image gallery, which are the same or similar with a given query text. Such a task is usually realized by matching a query text to the recognized words, outputted by an end-to-end scene text spotter. In this paper, we address this problem by directly learning a cross-modal similarity between a query text and each text instance from natural images. Specifically, we establish an end-to-end trainable network, jointly optimizing the procedures of scene text detection and cross-modal similarity learning. In this way, scene text retrieval can be simply performed by ranking the detected text instances with the learned similarity. Experiments on three benchmark datasets demonstrate our method consistently outperforms the state-of-the-art scene text spotting/retrieval approaches. In particular, the proposed framework of joint detection and similarity learning achieves significantly better performance than separated methods. Code is available at: https://github.com/lanfeng4659/STR-TDSL.

Scene text recognition is a challenging task due to the diversity in text appearance and complexity of natural scenes. Thanks to the development of deep learning and the large volume of training data, scene text recognition has made impressive progress in recent years. However, recent research on adversarial examples has shown that deep learning models are vulnerable to adversarial input with often imperceptible changes. As one of the most practical tasks in computer vision, scene text recognition is also facing huge security risks. To our best knowledge, there has been no work on adversarial attacks against scene text recognition. To investigate its effects on scene text recognition, we make the first attempt to attack the state-of-the-art scene text recognizers, i.e., attention-based recognizers. To that end, we first adjust the objective function designed for non-sequential tasks, such as image classification, semantic segmentation and image retrieval, to the sequential form. We then propose a novel and effective objective function to further reduce the amount of perturbation while achieving a higher attack success rate. Comprehensive experiments on several standard benchmarks clearly demonstrate effective adversarial effects on scene text recognition by the proposed attacks. It is note worthy that while this is harmful for a recognition system, it is highly advantageous for a text-based CAPTCHA system.

Scene text recognition (STR) is very challenging due to the diversity of text instances and the complexity of scenes. The community has paid increasing attention to boost the performance by improving the pre-processing image module, like rectification and deblurring, or the sequence translator. However, another critical module, i.e., the feature sequence extractor, has not been extensively explored. In this work, inspired by the success of neural architecture search (NAS), which can identify better architectures than human-designed ones, we propose automated STR (AutoSTR) to search data-dependent backbones to boost text recognition performance. First, we design a domain-specific search space for STR, which contains both choices on operations and constraints on the downsampling path. Then, we propose a two-step search algorithm, which decouples operations and downsampling path, for an efficient search in the given space. Experiments demonstrate that, by searching data-dependent backbones, AutoSTR can outperform the state-of-the-art approaches on standard benchmarks with much fewer FLOPS and model parameters.

Recently, end-to-end text spotting that aims to detect and recognize text from cluttered images simultaneously has received particularly growing interest in computer vision. Different from the existing approaches that formulate text detection as bounding box extraction or instance segmentation, we localize a set of points on the boundary of each text instance. With the representation of such boundary points, we establish a simple yet effective scheme for end-to-end text spotting, which can read the text of arbitrary shapes. Experiments on three challenging datasets, including ICDAR2015, TotalText and COCO-Text demonstrate that the proposed method consistently surpasses the state-of-the-art in both scene text detection and end-to-end text recognition tasks.

Person re-identification (Re-ID) requires discriminative features focusing on the full person to cope with inaccurate person bounding box detection, background clutter, and occlusion. Many recent person Re-ID methods attempt to learn such features describing full person details via part-based feature representation. However, the spatial context between these parts is ignored for the independent extractor on each separate part. In this paper, we propose to apply Long Short-Term Memory (LSTM) in an end-to-end way to model the pedestrian, seen as a sequence of body parts from head to foot. Integrating the contextual information strengthens the discriminative ability of local feature aligning better to full person. We also leverage the complementary information between local and global feature. Furthermore, we integrate both identification task and ranking task in one network, where a discriminative embedding and a similarity measurement are learned concurrently. This results in a novel three-branch framework named Deep-Person, which learns highly discriminative features for person Re-ID. Experimental results demonstrate that Deep-Person outperforms the state-of-the-art methods by a large margin on three challenging datasets including Market-1501, CUHK03, and DukeMTMC-reID.