Boqiang Zhang

Zixiao Wang, Hongtao Xie, Yuxin Wang et al.

In this paper, we explore the potential of the Contrastive Language-Image Pretraining (CLIP) model in scene text recognition (STR), and establish a novel Symmetrical Linguistic Feature Distillation framework (named CLIP-OCR) to leverage both visual and linguistic knowledge in CLIP. Different from previous CLIP-based methods mainly considering feature generalization on visual encoding, we propose a symmetrical distillation strategy (SDS) that further captures the linguistic knowledge in the CLIP text encoder. By cascading the CLIP image encoder with the reversed CLIP text encoder, a symmetrical structure is built with an image-to-text feature flow that covers not only visual but also linguistic information for distillation.Benefiting from the natural alignment in CLIP, such guidance flow provides a progressive optimization objective from vision to language, which can supervise the STR feature forwarding process layer-by-layer.Besides, a new Linguistic Consistency Loss (LCL) is proposed to enhance the linguistic capability by considering second-order statistics during the optimization. Overall, CLIP-OCR is the first to design a smooth transition between image and text for the STR task.Extensive experiments demonstrate the effectiveness of CLIP-OCR with 93.8% average accuracy on six popular STR benchmarks.Code will be available at https://github.com/wzx99/CLIPOCR.

Boqiang Zhang, Zuan Gao, Yadong Qu et al.

Visual text generation has significantly advanced through diffusion models aimed at producing images with readable and realistic text. Recent works primarily use a ControlNet-based framework, employing standard font text images to control diffusion models. Recognizing the critical role of control information in generating high-quality text, we investigate its influence from three perspectives: input encoding, role at different stages, and output features. Our findings reveal that: 1) Input control information has unique characteristics compared to conventional inputs like Canny edges and depth maps. 2) Control information plays distinct roles at different stages of the denoising process. 3) Output control features significantly differ from the base and skip features of the U-Net decoder in the frequency domain. Based on these insights, we propose TextGen, a novel framework designed to enhance generation quality by optimizing control information. We improve input and output features using Fourier analysis to emphasize relevant information and reduce noise. Additionally, we employ a two-stage generation framework to align the different roles of control information at different stages. Furthermore, we introduce an effective and lightweight dataset for training. Our method achieves state-of-the-art performance in both Chinese and English text generation. The code and dataset available at https://github.com/CyrilSterling/TextGen.

Bangbang Zhou, Yadong Qu, Zixiao Wang et al.

Recently, scene text recognition (STR) models have shown significant performance improvements. However, existing models still encounter difficulties in recognizing challenging texts that involve factors such as severely distorted and perspective characters. These challenging texts mainly cause two problems: (1) Large Intra-Class Variance. (2) Small Inter-Class Variance. An extremely distorted character may prominently differ visually from other characters within the same category, while the variance between characters from different classes is relatively small. To address the above issues, we propose a novel method that enriches the character features to enhance the discriminability of characters. Firstly, we propose the Character-Aware Constraint Encoder (CACE) with multiple blocks stacked. CACE introduces a decay matrix in each block to explicitly guide the attention region for each token. By continuously employing the decay matrix, CACE enables tokens to perceive morphological information at the character level. Secondly, an Intra-Inter Consistency Loss (I^2CL) is introduced to consider intra-class compactness and inter-class separability at feature space. I^2CL improves the discriminative capability of features by learning a long-term memory unit for each character category. Trained with synthetic data, our model achieves state-of-the-art performance on common benchmarks (94.1% accuracy) and Union14M-Benchmark (61.6% accuracy). Code is available at https://github.com/bang123-box/CFE.

Ronghao Dang, Yuqian Yuan, Wenqi Zhang et al.

The enhancement of generalization in robots by large vision-language models (LVLMs) is increasingly evident. Therefore, the embodied cognitive abilities of LVLMs based on egocentric videos are of great interest. However, current datasets for embodied video question answering lack comprehensive and systematic evaluation frameworks. Critical embodied cognitive issues, such as robotic self-cognition, dynamic scene perception, and hallucination, are rarely addressed. To tackle these challenges, we propose ECBench, a high-quality benchmark designed to systematically evaluate the embodied cognitive abilities of LVLMs. ECBench features a diverse range of scene video sources, open and varied question formats, and 30 dimensions of embodied cognition. To ensure quality, balance, and high visual dependence, ECBench uses class-independent meticulous human annotation and multi-round question screening strategies. Additionally, we introduce ECEval, a comprehensive evaluation system that ensures the fairness and rationality of the indicators. Utilizing ECBench, we conduct extensive evaluations of proprietary, open-source, and task-specific LVLMs. ECBench is pivotal in advancing the embodied cognitive capabilities of LVLMs, laying a solid foundation for developing reliable core models for embodied agents. All data and code are available at https://github.com/Rh-Dang/ECBench.

Boqiang Zhang, Lei Ke, Ruihan Yang et al.

Vision Language Model (VLM) development has largely relied on scaling model size, which hinders deployment on compute-constrained mobile and edge devices such as smartphones and robots. In this work, we explore the performance limits of compact (e.g., 2B and 8B) VLMs. We challenge the prevailing practice that state-of-the-art VLMs must rely on vision encoders initialized via massive contrastive pretraining (e.g., CLIP/SigLIP). We identify an objective mismatch: contrastive learning, optimized for discrimination, enforces coarse and category-level invariances that suppress fine-grained visual cues needed for dense captioning and complex VLM reasoning. To address this issue, we present Penguin-VL, whose vision encoder is initialized from a text-only LLM. Our experiments reveal that Penguin-Encoder serves as a superior alternative to traditional contrastive pretraining, unlocking a higher degree of visual fidelity and data efficiency for multimodal understanding. Across various image and video benchmarks, Penguin-VL achieves performance comparable to leading VLMs (e.g., Qwen3-VL) in mathematical reasoning and surpasses them in tasks such as document understanding, visual knowledge, and multi-perspective video understanding. Notably, these gains are achieved with a lightweight architecture, demonstrating that improved visual representation rather than model scaling is the primary driver of performance. Our ablations show that Penguin-Encoder consistently outperforms contrastive-pretrained encoders, preserving fine-grained spatial and temporal cues that are critical for dense perception and complex reasoning. This makes it a strong drop-in alternative for compute-efficient VLMs and enables high performance in resource-constrained settings. Code: https://github.com/tencent-ailab/Penguin-VL

Yuxin Wang, Lei Ke, Boqiang Zhang et al.

While current multimodal models can answer questions based on 2D images, they lack intrinsic 3D object perception, limiting their ability to comprehend spatial relationships and depth cues in 3D scenes. In this work, we propose N3D-VLM, a novel unified framework that seamlessly integrates native 3D object perception with 3D-aware visual reasoning, enabling both precise 3D grounding and interpretable spatial understanding. Unlike conventional end-to-end models that directly predict answers from RGB/RGB-D inputs, our approach equips the model with native 3D object perception capabilities, enabling it to directly localize objects in 3D space based on textual descriptions. Building upon accurate 3D object localization, the model further performs explicit reasoning in 3D, achieving more interpretable and structured spatial understanding. To support robust training for these capabilities, we develop a scalable data construction pipeline that leverages depth estimation to lift large-scale 2D annotations into 3D space, significantly increasing the diversity and coverage for 3D object grounding data, yielding over six times larger than the largest existing single-image 3D detection dataset. Moreover, the pipeline generates spatial question-answering datasets that target chain-of-thought (CoT) reasoning in 3D, facilitating joint training for both 3D object localization and 3D spatial reasoning. Experimental results demonstrate that our unified framework not only achieves state-of-the-art performance on 3D grounding tasks, but also consistently surpasses existing methods in 3D spatial reasoning in vision-language model.

Sicong Leng, Jing Wang, Jiaxi Li et al.

Large multimodal reasoning models have achieved rapid progress, but their advancement is constrained by two major limitations: the absence of open, large-scale, high-quality long chain-of-thought (CoT) data, and the instability of reinforcement learning (RL) algorithms in post-training. Group Relative Policy Optimization (GRPO), the standard framework for RL fine-tuning, is prone to gradient vanishing when reward variance is low, which weakens optimization signals and impairs convergence. This work makes three contributions: (1) We propose Variance-Aware Sampling (VAS), a data selection strategy guided by Variance Promotion Score (VPS) that combines outcome variance and trajectory diversity to promote reward variance and stabilize policy optimization. (2) We release large-scale, carefully curated resources containing ~1.6M long CoT cold-start data and ~15k RL QA pairs, designed to ensure quality, difficulty, and diversity, along with a fully reproducible end-to-end training codebase. (3) We open-source a family of multimodal reasoning models in multiple scales, establishing standardized baselines for the community. Experiments across mathematical reasoning benchmarks demonstrate the effectiveness of both the curated data and the proposed VAS. Comprehensive ablation studies and analyses provide further insight into the contributions of each component. In addition, we theoretically establish that reward variance lower-bounds the expected policy gradient magnitude, with VAS serving as a practical mechanism to realize this guarantee. Our code, data, and checkpoints are available at https://github.com/LengSicong/MMR1.

Boqiang Zhang, Kehan Li, Zesen Cheng et al. · pku

In this paper, we propose VideoLLaMA3, a more advanced multimodal foundation model for image and video understanding. The core design philosophy of VideoLLaMA3 is vision-centric. The meaning of "vision-centric" is two-fold: the vision-centric training paradigm and vision-centric framework design. The key insight of our vision-centric training paradigm is that high-quality image-text data is crucial for both image and video understanding. Instead of preparing massive video-text datasets, we focus on constructing large-scale and high-quality image-text datasets. VideoLLaMA3 has four training stages: 1) Vision Encoder Adaptation, which enables vision encoder to accept images of variable resolutions as input; 2) Vision-Language Alignment, which jointly tunes the vision encoder, projector, and LLM with large-scale image-text data covering multiple types (including scene images, documents, charts) as well as text-only data. 3) Multi-task Fine-tuning, which incorporates image-text SFT data for downstream tasks and video-text data to establish a foundation for video understanding. 4) Video-centric Fine-tuning, which further improves the model's capability in video understanding. As for the framework design, to better capture fine-grained details in images, the pretrained vision encoder is adapted to encode images of varying sizes into vision tokens with corresponding numbers, rather than a fixed number of tokens. For video inputs, we reduce the number of vision tokens according to their similarity so that the representation of videos will be more precise and compact. Benefit from vision-centric designs, VideoLLaMA3 achieves compelling performances in both image and video understanding benchmarks.

Guangqian Guo, Pengfei Chen, Yong Guo et al.

Segment Anything Model (SAM), known for its remarkable zero-shot segmentation capabilities, has garnered significant attention in the community. Nevertheless, its performance is challenged when dealing with what we refer to as visually non-salient scenarios, where there is low contrast between the foreground and background. In these cases, existing methods often cannot capture accurate contours and fail to produce promising segmentation results. In this paper, we propose Visually Non-Salient SAM (VNS-SAM), aiming to enhance SAM's perception of visually non-salient scenarios while preserving its original zero-shot generalizability. We achieve this by effectively exploiting SAM's low-level features through two designs: Mask-Edge Token Interactive decoder and Non-Salient Feature Mining module. These designs help the SAM decoder gain a deeper understanding of non-salient characteristics with only marginal parameter increments and computational requirements. The additional parameters of VNS-SAM can be optimized within 4 hours, demonstrating its feasibility and practicality. In terms of data, we established VNS-SEG, a unified dataset for various VNS scenarios, with more than 35K images, in contrast to previous single-task adaptations. It is designed to make the model learn more robust VNS features and comprehensively benchmark the model's segmentation performance and generalizability on VNS scenarios. Extensive experiments across various VNS segmentation tasks demonstrate the superior performance of VNS-SAM, particularly under zero-shot settings, highlighting its potential for broad real-world applications. Codes and datasets are publicly available at https://guangqian-guo.github.io/VNS-SAM.

Zuan Gao, Yuxin Wang, Yadong Qu et al.

In text recognition, self-supervised pre-training emerges as a good solution to reduce dependence on expansive annotated real data. Previous studies primarily focus on local visual representation by leveraging mask image modeling or sequence contrastive learning. However, they omit modeling the linguistic information in text images, which is crucial for recognizing text. To simultaneously capture local character features and linguistic information in visual space, we propose Symmetric Superimposition Modeling (SSM). The objective of SSM is to reconstruct the direction-specific pixel and feature signals from the symmetrically superimposed input. Specifically, we add the original image with its inverted views to create the symmetrically superimposed inputs. At the pixel level, we reconstruct the original and inverted images to capture character shapes and texture-level linguistic context. At the feature level, we reconstruct the feature of the same original image and inverted image with different augmentations to model the semantic-level linguistic context and the local character discrimination. In our design, we disrupt the character shape and linguistic rules. Consequently, the dual-level reconstruction facilitates understanding character shapes and linguistic information from the perspective of visual texture and feature semantics. Experiments on various text recognition benchmarks demonstrate the effectiveness and generality of SSM, with 4.1% average performance gains and 86.6% new state-of-the-art average word accuracy on Union14M benchmarks. The code is available at https://github.com/FaltingsA/SSM.

Boqiang Zhang, Hongtao Xie, Yuxin Wang et al.

Vision model have gained increasing attention due to their simplicity and efficiency in Scene Text Recognition (STR) task. However, due to lacking the perception of linguistic knowledge and information, recent vision models suffer from two problems: (1) the pure vision-based query results in attention drift, which usually causes poor recognition and is summarized as linguistic insensitive drift (LID) problem in this paper. (2) the visual feature is suboptimal for the recognition in some vision-missing cases (e.g. occlusion, etc.). To address these issues, we propose a $\textbf{L}$inguistic $\textbf{P}$erception $\textbf{V}$ision model (LPV), which explores the linguistic capability of vision model for accurate text recognition. To alleviate the LID problem, we introduce a Cascade Position Attention (CPA) mechanism that obtains high-quality and accurate attention maps through step-wise optimization and linguistic information mining. Furthermore, a Global Linguistic Reconstruction Module (GLRM) is proposed to improve the representation of visual features by perceiving the linguistic information in the visual space, which gradually converts visual features into semantically rich ones during the cascade process. Different from previous methods, our method obtains SOTA results while keeping low complexity (92.4% accuracy with only 8.11M parameters). Code is available at https://github.com/CyrilSterling/LPV.

Long Li, Weiwen Xu, Jiayan Guo et al. · pku

Effective research ideation is a critical step for scientific research. However, the exponential increase in scientific literature makes it challenging for researchers to stay current with recent advances and identify meaningful research directions. Recent developments in large language models~(LLMs) suggest a promising avenue for automating the generation of novel research ideas. However, existing methods for idea generation either trivially prompt LLMs or directly expose LLMs to extensive literature without indicating useful information. Inspired by the research process of human researchers, we propose a Chain-of-Ideas~(CoI) agent, an LLM-based agent that organizes relevant literature in a chain structure to effectively mirror the progressive development in a research domain. This organization facilitates LLMs to capture the current advancements in research, thereby enhancing their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation protocol that can comprehensively evaluate idea generation methods from different perspectives, aligning closely with the preferences of human researchers. Experimental results indicate that the CoI agent consistently outperforms other methods and shows comparable quality as humans in research idea generation. Moreover, our CoI agent is budget-friendly, with a minimum cost of \$0.50 to generate a candidate idea and its corresponding experimental design.

Yuqian Yuan, Hang Zhang, Wentong Li et al. · pku

Video Large Language Models (Video LLMs) have recently exhibited remarkable capabilities in general video understanding. However, they mainly focus on holistic comprehension and struggle with capturing fine-grained spatial and temporal details. Besides, the lack of high-quality object-level video instruction data and a comprehensive benchmark further hinders their advancements. To tackle these challenges, we introduce the VideoRefer Suite to empower Video LLM for finer-level spatial-temporal video understanding, i.e., enabling perception and reasoning on any objects throughout the video. Specially, we thoroughly develop VideoRefer Suite across three essential aspects: dataset, model, and benchmark. Firstly, we introduce a multi-agent data engine to meticulously curate a large-scale, high-quality object-level video instruction dataset, termed VideoRefer-700K. Next, we present the VideoRefer model, which equips a versatile spatial-temporal object encoder to capture precise regional and sequential representations. Finally, we meticulously create a VideoRefer-Bench to comprehensively assess the spatial-temporal understanding capability of a Video LLM, evaluating it across various aspects. Extensive experiments and analyses demonstrate that our VideoRefer model not only achieves promising performance on video referring benchmarks but also facilitates general video understanding capabilities.

Boqiang Zhang, Hongtao Xie, Zuan Gao et al.

Scene text images contain not only style information (font, background) but also content information (character, texture). Different scene text tasks need different information, but previous representation learning methods use tightly coupled features for all tasks, resulting in sub-optimal performance. We propose a Disentangled Representation Learning framework (DARLING) aimed at disentangling these two types of features for improved adaptability in better addressing various downstream tasks (choose what you really need). Specifically, we synthesize a dataset of image pairs with identical style but different content. Based on the dataset, we decouple the two types of features by the supervision design. Clearly, we directly split the visual representation into style and content features, the content features are supervised by a text recognition loss, while an alignment loss aligns the style features in the image pairs. Then, style features are employed in reconstructing the counterpart image via an image decoder with a prompt that indicates the counterpart's content. Such an operation effectively decouples the features based on their distinctive properties. To the best of our knowledge, this is the first time in the field of scene text that disentangles the inherent properties of the text images. Our method achieves state-of-the-art performance in Scene Text Recognition, Removal, and Editing.

Zhihang Liu, Chen-Wei Xie, Bin Wen et al.

Visual captioning benchmarks have become outdated with the emergence of modern multimodal large language models (MLLMs), as the brief ground-truth sentences and traditional metrics fail to assess detailed captions effectively. While recent benchmarks attempt to address this by focusing on keyword extraction or object-centric evaluation, they remain limited to vague-view or object-view analyses and incomplete visual element coverage. In this paper, we introduce CAPability, a comprehensive multi-view benchmark for evaluating visual captioning across 12 dimensions spanning six critical views. We curate nearly 11K human-annotated images and videos with visual element annotations to evaluate the generated captions. CAPability stably assesses both the correctness and thoroughness of captions with \textit{precision} and \textit{hit} metrics. By converting annotations to QA pairs, we further introduce a heuristic metric, \textit{know but cannot tell} ($K\bar{T}$), indicating a significant performance gap between QA and caption capabilities. Our work provides a holistic analysis of MLLMs' captioning abilities, as we identify their strengths and weaknesses across various dimensions, guiding future research to enhance specific aspects of their capabilities.