Abstract:The astonishing breakthrough of multimodal large language models (MLLMs) has necessitated new benchmarks to quantitatively assess their capabilities, reveal their limitations, and indicate future research directions. However, this is challenging in the context of remote sensing (RS), since the imagery features ultra-high resolution that incorporates extremely complex semantic relationships. Existing benchmarks usually adopt notably smaller image sizes than real-world RS scenarios, suffer from limited annotation quality, and consider insufficient dimensions of evaluation. To address these issues, we present XLRS-Bench: a comprehensive benchmark for evaluating the perception and reasoning capabilities of MLLMs in ultra-high-resolution RS scenarios. XLRS-Bench boasts the largest average image size (8500$\times$8500) observed thus far, with all evaluation samples meticulously annotated manually, assisted by a novel semi-automatic captioner on ultra-high-resolution RS images. On top of the XLRS-Bench, 16 sub-tasks are defined to evaluate MLLMs' 10 kinds of perceptual capabilities and 6 kinds of reasoning capabilities, with a primary emphasis on advanced cognitive processes that facilitate real-world decision-making and the capture of spatiotemporal changes. The results of both general and RS-focused MLLMs on XLRS-Bench indicate that further efforts are needed for real-world RS applications. We have open-sourced XLRS-Bench to support further research in developing more powerful MLLMs for remote sensing.
Abstract:Inspired by DeepSeek-R1's success in eliciting reasoning abilities through rule-based reinforcement learning (RL), we introduce Video-R1 as the first attempt to systematically explore the R1 paradigm for eliciting video reasoning within multimodal large language models (MLLMs). However, directly applying RL training with the GRPO algorithm to video reasoning presents two primary challenges: (i) a lack of temporal modeling for video reasoning, and (ii) the scarcity of high-quality video-reasoning data. To address these issues, we first propose the T-GRPO algorithm, which encourages models to utilize temporal information in videos for reasoning. Additionally, instead of relying solely on video data, we incorporate high-quality image-reasoning data into the training process. We have constructed two datasets: Video-R1-COT-165k for SFT cold start and Video-R1-260k for RL training, both comprising image and video data. Experimental results demonstrate that Video-R1 achieves significant improvements on video reasoning benchmarks such as VideoMMMU and VSI-Bench, as well as on general video benchmarks including MVBench and TempCompass, etc. Notably, Video-R1-7B attains a 35.8% accuracy on video spatial reasoning benchmark VSI-bench, surpassing the commercial proprietary model GPT-4o. All codes, models, data are released.
Abstract:Human experts excel at fine-grained visual discrimination by leveraging domain knowledge to refine perceptual features, a capability that remains underdeveloped in current Multimodal Large Language Models (MLLMs). Despite possessing vast expert-level knowledge, MLLMs struggle to integrate reasoning into visual perception, often generating direct responses without deeper analysis. To bridge this gap, we introduce knowledge-intensive visual grounding (KVG), a novel visual grounding task that requires both fine-grained perception and domain-specific knowledge integration. To address the challenges of KVG, we propose DeepPerception, an MLLM enhanced with cognitive visual perception capabilities. Our approach consists of (1) an automated data synthesis pipeline that generates high-quality, knowledge-aligned training samples, and (2) a two-stage training framework combining supervised fine-tuning for cognitive reasoning scaffolding and reinforcement learning to optimize perception-cognition synergy. To benchmark performance, we introduce KVG-Bench a comprehensive dataset spanning 10 domains with 1.3K manually curated test cases. Experimental results demonstrate that DeepPerception significantly outperforms direct fine-tuning, achieving +8.08\% accuracy improvements on KVG-Bench and exhibiting +4.60\% superior cross-domain generalization over baseline approaches. Our findings highlight the importance of integrating cognitive processes into MLLMs for human-like visual perception and open new directions for multimodal reasoning research. The data, codes, and models are released at https://github.com/thunlp/DeepPerception.
Abstract:Despite their impressive capabilities, Multimodal Large Language Models (MLLMs) face challenges with fine-grained perception and complex reasoning. Prevalent pre-training approaches focus on enhancing perception by training on high-quality image captions due to the extremely high cost of collecting chain-of-thought (CoT) reasoning data for improving reasoning. While leveraging advanced MLLMs for caption generation enhances scalability, the outputs often lack comprehensiveness and accuracy. In this paper, we introduce Self-Improving Cognition (SIcog), a self-learning framework designed to construct next-generation foundation MLLMs by enhancing their systematic cognitive capabilities through multimodal pre-training with self-generated data. Specifically, we propose chain-of-description, an approach that improves an MLLM's systematic perception by enabling step-by-step visual understanding, ensuring greater comprehensiveness and accuracy. Additionally, we adopt a structured CoT reasoning technique to enable MLLMs to integrate in-depth multimodal reasoning. To construct a next-generation foundation MLLM with self-improved cognition, SIcog first equips an MLLM with systematic perception and reasoning abilities using minimal external annotations. The enhanced models then generate detailed captions and CoT reasoning data, which are further curated through self-consistency. This curated data is ultimately used to refine the MLLM during multimodal pre-training, facilitating next-generation foundation MLLM construction. Extensive experiments on both low- and high-resolution MLLMs across diverse benchmarks demonstrate that, with merely 213K self-generated pre-training samples, SIcog produces next-generation foundation MLLMs with significantly improved cognition, achieving benchmark-leading performance compared to prevalent pre-training approaches.
Abstract:The recent advancement of Multimodal Large Language Models (MLLMs) has significantly improved their fine-grained perception of single images and general comprehension across multiple images. However, existing MLLMs still face challenges in achieving precise grounding in complex multi-image scenarios. To address this, we first explore a Chain-of-Thought (CoT) framework that integrates single-image grounding with multi-image comprehension. While partially effective, it remains unstable and struggles to capture abstract visual information due to its non-end-to-end nature. Therefore, we introduce Migician, the first multi-image grounding model capable of performing free-form and accurate grounding across multiple images. To support this, we present the MGrounding-630k dataset, which comprises data for several multi-image grounding tasks derived from existing datasets, along with newly generated free-form grounding instruction-following data. Furthermore, we propose MIG-Bench, a comprehensive benchmark specifically designed for evaluating multi-image grounding capabilities. Experimental results demonstrate that our model achieves significantly superior multi-image grounding capabilities, outperforming the best existing MLLMs by 21.61% and even surpassing much larger 70B models. Our code, model, dataset, and benchmark are fully open-sourced at https://migician-vg.github.io/.
Abstract:In multimodal large language models (MLLMs), vision transformers (ViTs) are widely employed for visual encoding. However, their performance in solving universal MLLM tasks is not satisfactory. We attribute it to a lack of information from diverse visual levels, impeding alignment with the various semantic granularity required for language generation. To address this issue, we present LLaVA-UHD v2, an advanced MLLM centered around a Hierarchical window transformer that enables capturing diverse visual granularity by constructing and integrating a high-resolution feature pyramid. As a vision-language projector, Hiwin transformer comprises two primary modules: (i) an inverse feature pyramid, constructed by a ViT-derived feature up-sampling process utilizing high-frequency details from an image pyramid, and (ii) hierarchical window attention, focusing on a set of key sampling features within cross-scale windows to condense multi-level feature maps. Extensive experiments demonstrate that LLaVA-UHD v2 achieves superior performance over existing MLLMs on popular benchmarks. Notably, our design brings an average boost of 3.7% across 14 benchmarks compared with the baseline method, 9.3% on DocVQA for instance. We make all the data, model checkpoint, and code publicly available to facilitate future research.
Abstract:In the current era of artificial intelligence, federated learning has emerged as a novel approach to addressing data privacy concerns inherent in centralized learning paradigms. This decentralized learning model not only mitigates the risk of data breaches but also enhances the system's scalability and robustness. However, this approach introduces a new challenge: how to fairly and accurately assess the contribution of each participant. Developing an effective contribution evaluation mechanism is crucial for federated learning. Such a mechanism incentivizes participants to actively contribute their data and computational resources, thereby improving the overall performance of the federated learning system. By allocating resources and rewards based on the size of the contributions, it ensures that each participant receives fair treatment, fostering sustained engagement.Currently, Shapley value-based methods are widely used to evaluate participants' contributions, with many researchers proposing modifications to adapt these methods to real-world scenarios. In this paper, we introduce a component called Dynamic Pruning Validation Set Shapley (DPVS-Shapley). This method accelerates the contribution assessment process by dynamically pruning the original dataset without compromising the evaluation's accuracy. Furthermore, this component can assign different weights to various samples, thereby allowing clients capable of distinguishing difficult examples to receive higher contribution scores.
Abstract:Masked Image Modeling (MIM) has emerged as a pivotal approach for developing foundational visual models in the field of remote sensing (RS). However, current RS datasets are limited in volume and diversity, which significantly constrains the capacity of MIM methods to learn generalizable representations. In this study, we introduce \textbf{RS-4M}, a large-scale dataset designed to enable highly efficient MIM training on RS images. RS-4M comprises 4 million optical images encompassing abundant and fine-grained RS visual tasks, including object-level detection and pixel-level segmentation. Compared to natural images, RS images often contain massive redundant background pixels, which limits the training efficiency of the conventional MIM models. To address this, we propose an efficient MIM method, termed \textbf{SelectiveMAE}, which dynamically encodes and reconstructs a subset of patch tokens selected based on their semantic richness. SelectiveMAE roots in a progressive semantic token selection module, which evolves from reconstructing semantically analogical tokens to encoding complementary semantic dependencies. This approach transforms conventional MIM training into a progressive feature learning process, enabling SelectiveMAE to efficiently learn robust representations of RS images. Extensive experiments show that SelectiveMAE significantly boosts training efficiency by 2.2-2.7 times and enhances the classification, detection, and segmentation performance of the baseline MIM model.The dataset, source code, and trained models will be released.
Abstract:Visual encoding constitutes the basis of large multimodal models (LMMs) in understanding the visual world. Conventional LMMs process images in fixed sizes and limited resolutions, while recent explorations in this direction are limited in adaptivity, efficiency, and even correctness. In this work, we first take GPT-4V and LLaVA-1.5 as representative examples and expose systematic flaws rooted in their visual encoding strategy. To address the challenges, we present LLaVA-UHD, a large multimodal model that can efficiently perceive images in any aspect ratio and high resolution. LLaVA-UHD includes three key components: (1) An image modularization strategy that divides native-resolution images into smaller variable-sized slices for efficient and extensible encoding, (2) a compression module that further condenses image tokens from visual encoders, and (3) a spatial schema to organize slice tokens for LLMs. Comprehensive experiments show that LLaVA-UHD outperforms established LMMs trained with 2-3 orders of magnitude more data on 9 benchmarks. Notably, our model built on LLaVA-1.5 336x336 supports 6 times larger (i.e., 672x1088) resolution images using only 94% inference computation, and achieves 6.4 accuracy improvement on TextVQA. Moreover, the model can be efficiently trained in academic settings, within 23 hours on 8 A100 GPUs (vs. 26 hours of LLaVA-1.5). We make the data and code publicly available at https://github.com/thunlp/LLaVA-UHD.
Abstract:In this paper, we propose a controllable dense captioner (ControlCap), which accommodates user's intention to dense captioning by introducing linguistic guidance. ControlCap is defined as a multimodal embedding bridging architecture, which comprises multimodal embedding generation (MEG) module and bi-directional embedding bridging (BEB) module. While MEG module represents objects/regions by combining embeddings of detailed information with context-aware ones, it also endows ControlCap the adaptability to specialized controls by utilizing them as linguistic guidance. BEB module aligns the linguistic guidance with visual embeddings through borrowing/returning features from/to the visual domain and gathering such features to predict text descriptions. Experiments on Visual Genome and VG-COCO datasets show that ControlCap respectively outperforms the state-of-the-art methods by 1.5% and 3.7% (mAP). Last but not least, with the capability of converting region-category pairs to region-text pairs, ControlCap is able to act as a powerful data engine for dense captioning. Code is available at https://github.com/callsys/ControlCap.