Qingpei Guo

Shiyu Xuan, Qingpei Guo, Ming Yang et al.

Multi-modal Large Language Models (MLLMs) have shown remarkable capabilities in various multi-modal tasks. Nevertheless, their performance in fine-grained image understanding tasks is still limited. To address this issue, this paper proposes a new framework to enhance the fine-grained image understanding abilities of MLLMs. Specifically, we present a new method for constructing the instruction tuning dataset at a low cost by leveraging annotations in existing datasets. A self-consistent bootstrapping method is also introduced to extend existing dense object annotations into high-quality referring-expression-bounding-box pairs. These methods enable the generation of high-quality instruction data which includes a wide range of fundamental abilities essential for fine-grained image perception. Moreover, we argue that the visual encoder should be tuned during instruction tuning to mitigate the gap between full image perception and fine-grained image perception. Experimental results demonstrate the superior performance of our method. For instance, our model exhibits a 5.2% accuracy improvement over Qwen-VL on GQA and surpasses the accuracy of Kosmos-2 by 24.7% on RefCOCO_val. We have also attained the top rank on the leaderboard of MMBench. This promising performance is achieved by training on only publicly available data, making it easily reproducible. The models, datasets, and codes are publicly available at https://github.com/SY-Xuan/Pink.

Chen Jiang, Hong Liu, Xuzheng Yu et al.

In recent years, the explosion of web videos makes text-video retrieval increasingly essential and popular for video filtering, recommendation, and search. Text-video retrieval aims to rank relevant text/video higher than irrelevant ones. The core of this task is to precisely measure the cross-modal similarity between texts and videos. Recently, contrastive learning methods have shown promising results for text-video retrieval, most of which focus on the construction of positive and negative pairs to learn text and video representations. Nevertheless, they do not pay enough attention to hard negative pairs and lack the ability to model different levels of semantic similarity. To address these two issues, this paper improves contrastive learning using two novel techniques. First, to exploit hard examples for robust discriminative power, we propose a novel Dual-Modal Attention-Enhanced Module (DMAE) to mine hard negative pairs from textual and visual clues. By further introducing a Negative-aware InfoNCE (NegNCE) loss, we are able to adaptively identify all these hard negatives and explicitly highlight their impacts in the training loss. Second, our work argues that triplet samples can better model fine-grained semantic similarity compared to pairwise samples. We thereby present a new Triplet Partial Margin Contrastive Learning (TPM-CL) module to construct partial order triplet samples by automatically generating fine-grained hard negatives for matched text-video pairs. The proposed TPM-CL designs an adaptive token masking strategy with cross-modal interaction to model subtle semantic differences. Extensive experiments demonstrate that the proposed approach outperforms existing methods on four widely-used text-video retrieval datasets, including MSR-VTT, MSVD, DiDeMo and ActivityNet.

Tian Gan, Xiao Wang, Yan Sun et al.

This paper aims to tackle a novel task - Temporal Sentence Grounding in Streaming Videos (TSGSV). The goal of TSGSV is to evaluate the relevance between a video stream and a given sentence query. Unlike regular videos, streaming videos are acquired continuously from a particular source, and are always desired to be processed on-the-fly in many applications such as surveillance and live-stream analysis. Thus, TSGSV is challenging since it requires the model to infer without future frames and process long historical frames effectively, which is untouched in the early methods. To specifically address the above challenges, we propose two novel methods: (1) a TwinNet structure that enables the model to learn about upcoming events; and (2) a language-guided feature compressor that eliminates redundant visual frames and reinforces the frames that are relevant to the query. We conduct extensive experiments using ActivityNet Captions, TACoS, and MAD datasets. The results demonstrate the superiority of our proposed methods. A systematic ablation study also confirms their effectiveness.

Qingpei Guo, Kaisheng Yao, Wei Chu

The ability to model intra-modal and inter-modal interactions is fundamental in multimodal machine learning. The current state-of-the-art models usually adopt deep learning models with fixed structures. They can achieve exceptional performances on specific tasks, but face a particularly challenging problem of modality mismatch because of diversity of input modalities and their fixed structures. In this paper, we present \textbf{Switch-BERT} for joint vision and language representation learning to address this problem. Switch-BERT extends BERT architecture by introducing learnable layer-wise and cross-layer interactions. It learns to optimize attention from a set of attention modes representing these interactions. One specific property of the model is that it learns to attend outputs from various depths, therefore mitigates the modality mismatch problem. We present extensive experiments on visual question answering, image-text retrieval and referring expression comprehension experiments. Results confirm that, whereas alternative architectures including ViLBERT and UNITER may excel in particular tasks, Switch-BERT can consistently achieve better or comparable performances than the current state-of-the-art models in these tasks. Ablation studies indicate that the proposed model achieves superior performances due to its ability in learning task-specific multimodal interactions.

Yutao Chen, Xingning Dong, Tian Gan et al.

Motivated by the superior performance of image diffusion models, more and more researchers strive to extend these models to the text-based video editing task. Nevertheless, current video editing tasks mainly suffer from the dilemma between the high fine-tuning cost and the limited generation capacity. Compared with images, we conjecture that videos necessitate more constraints to preserve the temporal consistency during editing. Towards this end, we propose EVE, a robust and efficient zero-shot video editing method. Under the guidance of depth maps and temporal consistency constraints, EVE derives satisfactory video editing results with an affordable computational and time cost. Moreover, recognizing the absence of a publicly available video editing dataset for fair comparisons, we construct a new benchmark ZVE-50 dataset. Through comprehensive experimentation, we validate that EVE could achieve a satisfactory trade-off between performance and efficiency. We will release our dataset and codebase to facilitate future researchers.

Ruobing Zheng, Tianqi Li, Jianing Li et al.

Reasoning post-training improves Large Language Models (LLMs) on complex tasks such as mathematics and coding, but its benefits across diverse multimodal tasks remains uncertain. The trend of releasing parallel "Instruct" and "Thinking" models by leading teams is both resource-intensive and user-unfriendly. Prior work finds that the gains from reasoning training are influenced by multiple factors, such as base model capabilities, task characteristics, and Chain-of-Thought (CoT) data quality. However, principled criteria for determining when reasoning post-training is beneficial and which data should support it are still lacking. In this paper, we propose Dual Tuning, a reasoning efficacy-driven data curation framework for multimodal LLMs training. Given a target task and a base model, Dual Tuning jointly evaluates whether the training data is beneficial and whether reasoning training with current CoT content yields positive gains over non-reasoning alternatives. We apply Dual Tuning across spatial, mathematical, and multi-disciplinary tasks, and further analyze how reinforcement learning and thinking patterns affect reasoning efficacy. The Dual Tuning results guide data curation by identifying data that benefit reasoning training, data better suited to direct-answer training, and data that are detrimental under both training modes. Our work provides quantitative criteria for selecting appropriate training data and matching post-training strategies.

Yuyan Chen, Yiwen Qian, Songzhou Yan et al.

In the era of social media video platforms, popular ``hot-comments'' play a crucial role in attracting user impressions of short-form videos, making them vital for marketing and branding purpose. However, existing research predominantly focuses on generating descriptive comments or ``danmaku'' in English, offering immediate reactions to specific video moments. Addressing this gap, our study introduces \textsc{HotVCom}, the largest Chinese video hot-comment dataset, comprising 94k diverse videos and 137 million comments. We also present the \texttt{ComHeat} framework, which synergistically integrates visual, auditory, and textual data to generate influential hot-comments on the Chinese video dataset. Empirical evaluations highlight the effectiveness of our framework, demonstrating its excellence on both the newly constructed and existing datasets.

Harry Cheng, Yangyang Guo, Qingpei Guo et al.

Multi-modal Large Language Models (MLLMs) have dramatically advanced the research field and delivered powerful vision-language understanding capabilities. However, these models often inherit deep-rooted social biases from their training data, leading to uncomfortable responses with respect to attributes such as race and gender. This paper addresses the issue of social biases in MLLMs by i) introducing a comprehensive counterfactual dataset with multiple social concepts (CMSC), which complements existing datasets by providing 18 diverse and balanced social concepts; and ii) proposing a counter-stereotype debiasing (CSD) strategy that mitigates social biases in MLLMs by leveraging the opposites of prevalent stereotypes. CSD incorporates both a novel bias-aware data sampling method and a loss rescaling method, enabling the model to effectively reduce biases. We conduct extensive experiments with four prevalent MLLM architectures. The results demonstrate the advantage of the CMSC dataset and the edge of CSD strategy in reducing social biases compared to existing competing methods, without compromising the overall performance on general multi-modal reasoning benchmarks.

Inclusion AI, Biao Gong, Cheng Zou et al.

We propose Ming-Omni, a unified multimodal model capable of processing images, text, audio, and video, while demonstrating strong proficiency in both speech and image generation. Ming-Omni employs dedicated encoders to extract tokens from different modalities, which are then processed by Ling, an MoE architecture equipped with newly proposed modality-specific routers. This design enables a single model to efficiently process and fuse multimodal inputs within a unified framework, thereby facilitating diverse tasks without requiring separate models, task-specific fine-tuning, or structural redesign. Importantly, Ming-Omni extends beyond conventional multimodal models by supporting audio and image generation. This is achieved through the integration of an advanced audio decoder for natural-sounding speech and Ming-Lite-Uni for high-quality image generation, which also allow the model to engage in context-aware chatting, perform text-to-speech conversion, and conduct versatile image editing. Our experimental results showcase Ming-Omni offers a powerful solution for unified perception and generation across all modalities. Notably, our proposed Ming-Omni is the first open-source model we are aware of to match GPT-4o in modality support, and we release all code and model weights to encourage further research and development in the community.

Qingpei Guo, Kaiyou Song, Zipeng Feng et al.

We present M2-omni, a cutting-edge, open-source omni-MLLM that achieves competitive performance to GPT-4o. M2-omni employs a unified multimodal sequence modeling framework, which empowers Large Language Models(LLMs) to acquire comprehensive cross-modal understanding and generation capabilities. Specifically, M2-omni can process arbitrary combinations of audio, video, image, and text modalities as input, generating multimodal sequences interleaving with audio, image, or text outputs, thereby enabling an advanced and interactive real-time experience. The training of such an omni-MLLM is challenged by significant disparities in data quantity and convergence rates across modalities. To address these challenges, we propose a step balance strategy during pre-training to handle the quantity disparities in modality-specific data. Additionally, a dynamically adaptive balance strategy is introduced during the instruction tuning stage to synchronize the modality-wise training progress, ensuring optimal convergence. Notably, we prioritize preserving strong performance on pure text tasks to maintain the robustness of M2-omni's language understanding capability throughout the training process. To our best knowledge, M2-omni is currently a very competitive open-source model to GPT-4o, characterized by its comprehensive modality and task support, as well as its exceptional performance. We expect M2-omni will advance the development of omni-MLLMs, thus facilitating future research in this domain.

Inclusion AI, Biao Gong, Cheng Zou et al.

We introduce Ming-Lite-Uni, an open-source multimodal framework featuring a newly designed unified visual generator and a native multimodal autoregressive model tailored for unifying vision and language. Specifically, this project provides an open-source implementation of the integrated MetaQueries and M2-omni framework, while introducing the novel multi-scale learnable tokens and multi-scale representation alignment strategy. By leveraging a fixed MLLM and a learnable diffusion model, Ming-Lite-Uni enables native multimodal AR models to perform both text-to-image generation and instruction based image editing tasks, expanding their capabilities beyond pure visual understanding. Our experimental results demonstrate the strong performance of Ming-Lite-Uni and illustrate the impressive fluid nature of its interactive process. All code and model weights are open-sourced to foster further exploration within the community. Notably, this work aligns with concurrent multimodal AI milestones - such as ChatGPT-4o with native image generation updated in March 25, 2025 - underscoring the broader significance of unified models like Ming-Lite-Uni on the path toward AGI. Ming-Lite-Uni is in alpha stage and will soon be further refined.

Xingning Dong, Qingpei Guo, Tian Gan et al.

We present a framework for learning cross-modal video representations by directly pre-training on raw data to facilitate various downstream video-text tasks. Our main contributions lie in the pre-training framework and proxy tasks. First, based on the shortcomings of two mainstream pixel-level pre-training architectures (limited applications or less efficient), we propose Shared Network Pre-training (SNP). By employing one shared BERT-type network to refine textual and cross-modal features simultaneously, SNP is lightweight and could support various downstream applications. Second, based on the intuition that people always pay attention to several "significant words" when understanding a sentence, we propose the Significant Semantic Strengthening (S3) strategy, which includes a novel masking and matching proxy task to promote the pre-training performance. Experiments conducted on three downstream video-text tasks and six datasets demonstrate that, we establish a new state-of-the-art in pixel-level video-text pre-training; we also achieve a satisfactory balance between the pre-training efficiency and the fine-tuning performance. The codebase are available at https://github.com/alipay/Ant-Multi-Modal-Framework/tree/main/prj/snps3_vtp.

Xingning Dong, Zipeng Feng, Chunluan Zhou et al.

We present a Multi-Modal Recipe for Advancing Adaptation-based Pre-training towards effective and efficient zero-shot video-text retrieval, dubbed M2-RAAP. Upon popular image-text models like CLIP, most current adaptation-based video-text pre-training methods are confronted by three major issues, i.e., noisy data corpus, time-consuming pre-training, and limited performance gain. Towards this end, we conduct a comprehensive study including four critical steps in video-text pre-training. Specifically, we investigate 1) data filtering and refinement, 2) video input type selection, 3) temporal modeling, and 4) video feature enhancement. We then summarize this empirical study into the M2-RAAP recipe, where our technical contributions lie in 1) the data filtering and text re-writing pipeline resulting in 1M high-quality bilingual video-text pairs, 2) the replacement of video inputs with key-frames to accelerate pre-training, and 3) the Auxiliary-Caption-Guided (ACG) strategy to enhance video features. We conduct extensive experiments by adapting three image-text foundation models on two refined video-text datasets from different languages, validating the robustness and reproducibility of M2-RAAP for adaptation-based pre-training. Results demonstrate that M2-RAAP yields superior performance with significantly reduced data (-90%) and time consumption (-95%), establishing a new SOTA on four English zero-shot retrieval datasets and two Chinese ones. We are preparing our refined bilingual data annotations and codebase, which will be available at https://github.com/alipay/Ant-Multi-Modal-Framework/tree/main/prj/M2_RAAP.

Ruihan Xu, Qingpei Guo, Yao Zhu et al.

Recent works have indicated redundancy across transformer blocks, prompting the research of depth compression to prune less crucial blocks. However, current ways of entire-block pruning suffer from risks of discarding meaningful cues learned in those blocks, leading to substantial performance degradation. As another line of model compression, channel pruning can better preserve performance, while it cannot reduce model depth and is challenged by inconsistent pruning ratios for individual layers. To pursue better model compression and acceleration, this paper proposes \textbf{FlattenGPT}, a novel way to detect and reduce depth-wise redundancies. By flatting two adjacent blocks into one, it compresses the network depth, meanwhile enables more effective parameter redundancy detection and removal. FlattenGPT allows to preserve the knowledge learned in all blocks, and remains consistent with the original transformer architecture. Extensive experiments demonstrate that FlattenGPT enhances model efficiency with a decent trade-off to performance. It outperforms existing pruning methods in both zero-shot accuracies and WikiText-2 perplexity across various model types and parameter sizes. On LLaMA-2/3 and Qwen-1.5 models, FlattenGPT retains 90-96\% of zero-shot performance with a compression ratio of 20\%. It also outperforms other pruning methods in accelerating LLM inference, making it promising for enhancing the efficiency of transformers.

Hengyuan Zhao, Ziqin Wang, Qixin Sun et al.

Mixture of Experts (MoE) architectures have recently advanced the scalability and adaptability of large language models (LLMs) for continual multimodal learning. However, efficiently extending these models to accommodate sequential tasks remains challenging. As new tasks arrive, naive model expansion leads to rapid parameter growth, while modifying shared routing components often causes catastrophic forgetting, undermining previously learned knowledge. To address these issues, we propose LLaVA-CMoE, a continual learning framework for LLMs that requires no replay data of previous tasks and ensures both parameter efficiency and robust knowledge retention. Our approach introduces a Probe-Guided Knowledge Extension mechanism, which uses probe experts to dynamically determine when and where new experts should be added, enabling adaptive and minimal parameter expansion tailored to task complexity. Furthermore, we present a Probabilistic Task Locator that assigns each task a dedicated, lightweight router. To handle the practical issue that task labels are unknown during inference, we leverage a VAE-based reconstruction strategy to identify the most suitable router by matching input distributions, allowing automatic and accurate expert allocation. This design mitigates routing conflicts and catastrophic forgetting, enabling robust continual learning without explicit task labels. Extensive experiments on the CoIN benchmark, covering eight diverse VQA tasks, demonstrate that LLaVA-CMoE delivers strong continual learning performance with a compact model size, significantly reducing forgetting and parameter overhead compared to prior methods. These results showcase the effectiveness and scalability of our approach for parameter-efficient continual learning in large language models. Our code will be open-sourced soon.

Tan Pan, Furong Xu, Xudong Yang et al.

Image retrieval plays an important role in the Internet world. Usually, the core parts of mainstream visual retrieval systems include an online service of the embedding model and a large-scale vector database. For traditional model upgrades, the old model will not be replaced by the new one until the embeddings of all the images in the database are re-computed by the new model, which takes days or weeks for a large amount of data. Recently, backward-compatible training (BCT) enables the new model to be immediately deployed online by making the new embeddings directly comparable to the old ones. For BCT, improving the compatibility of two models with less negative impact on retrieval performance is the key challenge. In this paper, we introduce AdvBCT, an Adversarial Backward-Compatible Training method with an elastic boundary constraint that takes both compatibility and discrimination into consideration. We first employ adversarial learning to minimize the distribution disparity between embeddings of the new model and the old model. Meanwhile, we add an elastic boundary constraint during training to improve compatibility and discrimination efficiently. Extensive experiments on GLDv2, Revisited Oxford (ROxford), and Revisited Paris (RParis) demonstrate that our method outperforms other BCT methods on both compatibility and discrimination. The implementation of AdvBCT will be publicly available at https://github.com/Ashespt/AdvBCT.

Li Jiang, Yusen Wu, Junwu Xiong et al.

Preference datasets are essential for incorporating human preferences into pre-trained language models, playing a key role in the success of Reinforcement Learning from Human Feedback. However, these datasets often demonstrate conflicting alignment objectives, leading to increased vulnerability to jailbreak attacks and challenges in adapting downstream tasks to prioritize specific alignment objectives without negatively impacting others. In this work, we introduce a novel statistical metric, Alignment Dimension Conflict, to quantify the degree of conflict within preference datasets. We then present \texttt{Hummer} and its fine-grained variant, \texttt{Hummer-F}, as innovative pairwise preference datasets with reduced-conflict alignment objectives. \texttt{Hummer} is built based on UltraFeedback and is enhanced by AI feedback from GPT-4, marking as the first preference dataset aimed at reducing the competition between alignment objectives. Furthermore, we develop reward models, HummerRM and HummerRM-F, which employ a hybrid sampling approach to balance diverse alignment objectives effectively. This sampling method positions HummerRM as an ideal model for domain-specific further fine-tuning and reducing vulnerabilities to attacks.

Inclusion AI, Fudong Wang, Jiajia Liu et al.

Recent advancements in Multimodal Large Language Models (MLLMs), particularly through Reinforcement Learning with Verifiable Rewards (RLVR), have significantly enhanced their reasoning abilities. However, a critical gap persists: these models struggle with dynamic spatial interactions, a capability essential for real-world applications. To bridge this gap, we introduce M2-Reasoning-7B, a model designed to excel in both general and spatial reasoning. Our approach integrates two key innovations: (1) a novel data pipeline that generates 294.2K high-quality data samples (168K for cold-start fine-tuning and 126.2K for RLVR), which feature logically coherent reasoning trajectories and have undergone comprehensive assessment; and (2) a dynamic multi-task training strategy with step-wise optimization to mitigate conflicts between data, and task-specific rewards for delivering tailored incentive signals. This combination of curated data and advanced training allows M2-Reasoning-7B to set a new state-of-the-art (SOTA) across 8 benchmarks, showcasing superior performance in both general and spatial reasoning domains.

Muzhi Zhu, Yuzhuo Tian, Hao Chen et al.

While MLLMs have demonstrated adequate image understanding capabilities, they still struggle with pixel-level comprehension, limiting their practical applications. Current evaluation tasks like VQA and visual grounding remain too coarse to assess fine-grained pixel comprehension accurately. Though segmentation is foundational for pixel-level understanding, existing methods often require MLLMs to generate implicit tokens, decoded through external pixel decoders. This approach disrupts the MLLM's text output space, potentially compromising language capabilities and reducing flexibility and extensibility, while failing to reflect the model's intrinsic pixel-level understanding. Thus, we introduce the Human-Like Mask Annotation Task (HLMAT), a new paradigm where MLLMs mimic human annotators using interactive segmentation tools. Modeling segmentation as a multi-step Markov Decision Process, HLMAT enables MLLMs to iteratively generate text-based click points, achieving high-quality masks without architectural changes or implicit tokens. Through this setup, we develop SegAgent, a model fine-tuned on human-like annotation trajectories, which achieves performance comparable to state-of-the-art (SOTA) methods and supports additional tasks like mask refinement and annotation filtering. HLMAT provides a protocol for assessing fine-grained pixel understanding in MLLMs and introduces a vision-centric, multi-step decision-making task that facilitates exploration of MLLMs' visual reasoning abilities. Our adaptations of policy improvement method StaR and PRM-guided tree search further enhance model robustness in complex segmentation tasks, laying a foundation for future advancements in fine-grained visual perception and multi-step decision-making for MLLMs.

Ziping Ma, Furong Xu, Jian Liu et al.

Multimodal alignment between language and vision is the fundamental topic in current vision-language model research. Contrastive Captioners (CoCa), as a representative method, integrates Contrastive Language-Image Pretraining (CLIP) and Image Caption (IC) into a unified framework, resulting in impressive results. CLIP imposes a bidirectional constraints on global representation of entire images and sentences. Although IC conducts an unidirectional image-to-text generation on local representation, it lacks any constraint on local text-to-image reconstruction, which limits the ability to understand images at a fine-grained level when aligned with texts. To achieve multimodal alignment from both global and local perspectives, this paper proposes Symmetrizing Contrastive Captioners (SyCoCa), which introduces bidirectional interactions on images and texts across the global and local representation levels. Specifically, we expand a Text-Guided Masked Image Modeling (TG-MIM) head based on ITC and IC heads. The improved SyCoCa can further leverage textual cues to reconstruct contextual images and visual cues to predict textual contents. When implementing bidirectional local interactions, the local contents of images tend to be cluttered or unrelated to their textual descriptions. Thus, we employ an attentive masking strategy to select effective image patches for interaction. Extensive experiments on five vision-language tasks, including image-text retrieval, image-captioning, visual question answering, and zero-shot/finetuned image classification, validate the effectiveness of our proposed method.

Qingpei Guo, Furong Xu, Hanxiao Zhang et al.

Vision-language foundation models like CLIP have revolutionized the field of artificial intelligence. Nevertheless, VLM models supporting multi-language, e.g., in both Chinese and English, have lagged due to the relative scarcity of large-scale pretraining datasets. Toward this end, we introduce a comprehensive bilingual (Chinese-English) dataset BM-6B with over 6 billion image-text pairs, aimed at enhancing multimodal foundation models to well understand images in both languages. To handle such a scale of dataset, we propose a novel grouped aggregation approach for image-text contrastive loss computation, which reduces the communication overhead and GPU memory demands significantly, facilitating a 60% increase in training speed. We pretrain a series of bilingual image-text foundation models with an enhanced fine-grained understanding ability on BM-6B, the resulting models, dubbed as $M^2$-Encoders (pronounced "M-Square"), set new benchmarks in both languages for multimodal retrieval and classification tasks. Notably, Our largest $M^2$-Encoder-10B model has achieved top-1 accuracies of 88.5% on ImageNet and 80.7% on ImageNet-CN under a zero-shot classification setting, surpassing previously reported SoTA methods by 2.2% and 21.1%, respectively. The $M^2$-Encoder series represents one of the most comprehensive bilingual image-text foundation models to date, so we are making it available to the research community for further exploration and development.

Ziyuan Huang, DanDan Zheng, Cheng Zou et al.

Visual tokenization remains a core challenge in unifying visual understanding and generation within the autoregressive paradigm. Existing methods typically employ tokenizers in discrete latent spaces to align with the tokens from large language models, where the quantization errors can limit semantic expressiveness and degrade the capability of vision-language understanding. To address this, we introduce MingTok, a new family of visual tokenizers with a continuous latent space, for unified autoregressive generation and understanding. While understanding tasks favor discriminative high-dimensional features, generation tasks prefer compact low-level codes. Thus, to reconcile these competing demands, MingTok adopts a three-stage sequential architecture involving low-level encoding, semantic expansion, and visual reconstruction. Built on top of it, Ming-UniVision eliminates the need for task-specific visual representations, and unifies diverse vision-language tasks under a single autoregrsssive prediction paradigm. By formulating both understanding and generation as next-token prediction in a shared continuous space, it seamlessly supports multi-round, in-context tasks such as iterative understanding, generation and editing. Empirically, we find that using a unified continuous visual representation reconciles the competing requirements on the tokenizers by the understanding and generation tasks, thereby leading to state-of-the-art level performance across both domains. We hope our findings will facilitate unified visual tokenization in the continuous domain. Inference code and model weights are released to benefit community.

Xuzheng Yu, Chen Jiang, Xingning Dong et al.

The user base of short video apps has experienced unprecedented growth in recent years, resulting in a significant demand for video content analysis. In particular, text-video retrieval, which aims to find the top matching videos given text descriptions from a vast video corpus, is an essential function, the primary challenge of which is to bridge the modality gap. Nevertheless, most existing approaches treat texts merely as discrete tokens and neglect their syntax structures. Moreover, the abundant spatial and temporal clues in videos are often underutilized due to the lack of interaction with text. To address these issues, we argue that using texts as guidance to focus on relevant temporal frames and spatial regions within videos is beneficial. In this paper, we propose a novel Syntax-Hierarchy-Enhanced text-video retrieval method (SHE-Net) that exploits the inherent semantic and syntax hierarchy of texts to bridge the modality gap from two perspectives. First, to facilitate a more fine-grained integration of visual content, we employ the text syntax hierarchy, which reveals the grammatical structure of text descriptions, to guide the visual representations. Second, to further enhance the multi-modal interaction and alignment, we also utilize the syntax hierarchy to guide the similarity calculation. We evaluated our method on four public text-video retrieval datasets of MSR-VTT, MSVD, DiDeMo, and ActivityNet. The experimental results and ablation studies confirm the advantages of our proposed method.

Yudong Han, Qingpei Guo, Liyuan Pan et al.

The challenge in LLM-based video understanding lies in preserving visual and semantic information in long videos while maintaining a memory-affordable token count. However, redundancy and correspondence in videos have hindered the performance potential of existing methods. Through statistical learning on current datasets, we observe that redundancy occurs in both repeated and answer-irrelevant frames, and the corresponding frames vary with different questions. This suggests the possibility of adopting dynamic encoding to balance detailed video information preservation with token budget reduction. To this end, we propose a dynamic cooperative network, DynFocus, for memory-efficient video encoding in this paper. Specifically, i) a Dynamic Event Prototype Estimation (DPE) module to dynamically select meaningful frames for question answering; (ii) a Compact Cooperative Encoding (CCE) module that encodes meaningful frames with detailed visual appearance and the remaining frames with sketchy perception separately. We evaluate our method on five publicly available benchmarks, and experimental results consistently demonstrate that our method achieves competitive performance.

Yabing Wang, Zhuotao Tian, Qingpei Guo et al.

Composed Image Retrieval (CIR) is a challenging multimodal task that retrieves a target image based on a reference image and accompanying modification text. Due to the high cost of annotating CIR triplet datasets, zero-shot (ZS) CIR has gained traction as a promising alternative. Existing studies mainly focus on projection-based methods, which map an image to a single pseudo-word token. However, these methods face three critical challenges: (1) insufficient pseudo-word token representation capacity, (2) discrepancies between training and inference phases, and (3) reliance on large-scale synthetic data. To address these issues, we propose a two-stage framework where the training is accomplished from mapping to composing. In the first stage, we enhance image-to-pseudo-word token learning by introducing a visual semantic injection module and a soft text alignment objective, enabling the token to capture richer and fine-grained image information. In the second stage, we optimize the text encoder using a small amount of synthetic triplet data, enabling it to effectively extract compositional semantics by combining pseudo-word tokens with modification text for accurate target image retrieval. The strong visual-to-pseudo mapping established in the first stage provides a solid foundation for the second stage, making our approach compatible with both high- and low-quality synthetic data, and capable of achieving significant performance gains with only a small amount of synthetic data. Extensive experiments were conducted on three public datasets, achieving superior performance compared to existing approaches.

Yabing Wang, Zhuotao Tian, Qingpei Guo et al.

Visual Grounding aims to localize the referring object in an image given a natural language expression. Recent advancements in DETR-based visual grounding methods have attracted considerable attention, as they directly predict the coordinates of the target object without relying on additional efforts, such as pre-generated proposal candidates or pre-defined anchor boxes. However, existing research primarily focuses on designing stronger multi-modal decoder, which typically generates learnable queries by random initialization or by using linguistic embeddings. This vanilla query generation approach inevitably increases the learning difficulty for the model, as it does not involve any target-related information at the beginning of decoding. Furthermore, they only use the deepest image feature during the query learning process, overlooking the importance of features from other levels. To address these issues, we propose a novel approach, called RefFormer. It consists of the query adaption module that can be seamlessly integrated into CLIP and generate the referential query to provide the prior context for decoder, along with a task-specific decoder. By incorporating the referential query into the decoder, we can effectively mitigate the learning difficulty of the decoder, and accurately concentrate on the target object. Additionally, our proposed query adaption module can also act as an adapter, preserving the rich knowledge within CLIP without the need to tune the parameters of the backbone network. Extensive experiments demonstrate the effectiveness and efficiency of our proposed method, outperforming state-of-the-art approaches on five visual grounding benchmarks.

Xuelin Zhu, Jiuxin Cao, Jian liu et al.

Pre-trained vision-language models have notably accelerated progress of open-world concept recognition. Their impressive zero-shot ability has recently been transferred to multi-label image classification via prompt tuning, enabling to discover novel labels in an open-vocabulary manner. However, this paradigm suffers from non-trivial training costs, and becomes computationally prohibitive for a large number of candidate labels. To address this issue, we note that vision-language pre-training aligns images and texts in a unified embedding space, making it potential for an adapter network to identify labels in visual modality while be trained in text modality. To enhance such cross-modal transfer ability, a simple yet effective method termed random perturbation is proposed, which enables the adapter to search for potential visual embeddings by perturbing text embeddings with noise during training, resulting in better performance in visual modality. Furthermore, we introduce an effective approach to employ large language models for multi-label instruction-following text generation. In this way, a fully automated pipeline for visual label recognition is developed without relying on any manual data. Extensive experiments on public benchmarks show the superiority of our method in various multi-label classification tasks.

Inclusion AI, Bowen Ma, Cheng Zou et al.

We propose Ming-Flash-Omni, an upgraded version of Ming-Omni, built upon a sparser Mixture-of-Experts (MoE) variant of Ling-Flash-2.0 with 100 billion total parameters, of which only 6.1 billion are active per token. This architecture enables highly efficient scaling (dramatically improving computational efficiency while significantly expanding model capacity) and empowers stronger unified multimodal intelligence across vision, speech, and language, representing a key step toward Artificial General Intelligence (AGI). Compared to its predecessor, the upgraded version exhibits substantial improvements across multimodal understanding and generation. We significantly advance speech recognition capabilities, achieving state-of-the-art performance in contextual ASR and highly competitive results in dialect-aware ASR. In image generation, Ming-Flash-Omni introduces high-fidelity text rendering and demonstrates marked gains in scene consistency and identity preservation during image editing. Furthermore, Ming-Flash-Omni introduces generative segmentation, a capability that not only achieves strong standalone segmentation performance but also enhances spatial control in image generation and improves editing consistency. Notably, Ming-Flash-Omni achieves state-of-the-art results in text-to-image generation and generative segmentation, and sets new records on all 12 contextual ASR benchmarks, all within a single unified architecture.

Linglin Jing, Yuting Gao, Zhigang Wang et al.

Recent advancements have shown that the Mixture of Experts (MoE) approach significantly enhances the capacity of large language models (LLMs) and improves performance on downstream tasks. Building on these promising results, multi-modal large language models (MLLMs) have increasingly adopted MoE techniques. However, existing multi-modal MoE tuning methods typically face two key challenges: expert uniformity and router rigidity. Expert uniformity occurs because MoE experts are often initialized by simply replicating the FFN parameters from LLMs, leading to homogenized expert functions and weakening the intended diversification of the MoE architecture. Meanwhile, router rigidity stems from the prevalent use of static linear routers for expert selection, which fail to distinguish between visual and textual tokens, resulting in similar expert distributions for image and text. To address these limitations, we propose EvoMoE, an innovative MoE tuning framework. EvoMoE introduces a meticulously designed expert initialization strategy that progressively evolves multiple robust experts from a single trainable expert, a process termed expert evolution that specifically targets severe expert homogenization. Furthermore, we introduce the Dynamic Token-aware Router (DTR), a novel routing mechanism that allocates input tokens to appropriate experts based on their modality and intrinsic token values. This dynamic routing is facilitated by hypernetworks, which dynamically generate routing weights tailored for each individual token. Extensive experiments demonstrate that EvoMoE significantly outperforms other sparse MLLMs across a variety of multi-modal benchmarks, including MME, MMBench, TextVQA, and POPE. Our results highlight the effectiveness of EvoMoE in enhancing the performance of MLLMs by addressing the critical issues of expert uniformity and router rigidity.

Ruihan Xu, Yuting Gao, Lan Wang et al.

Large Multimodal Models (LMMs) have achieved remarkable success in vision-language tasks, yet their vast parameter counts are often underutilized during both training and inference. In this work, we embrace the idea of looping back to move forward: reusing model parameters through recursive refinement to extract stronger multimodal representations without increasing model size. We propose RecursiveVLM, a recursive Transformer architecture tailored for LMMs. Two key innovations enable effective looping: (i) a Recursive Connector that aligns features across recursion steps by fusing intermediate-layer hidden states and applying modality-specific projections, respecting the distinct statistical structures of vision and language tokens; (ii) a Monotonic Recursion Loss that supervises every step and guarantees performance improves monotonically with recursion depth. This design transforms recursion into an on-demand refinement mechanism: delivering strong results with few loops on resource-constrained devices and progressively improving outputs when more computation resources are available. Experiments show consistent gains of +3% over standard Transformers and +7% over vanilla recursive baselines, demonstrating that strategic looping is a powerful path toward efficient, deployment-adaptive LMMs.

Yuting Gao, Weihao Chen, Lan Wang et al.

Preference learning has recently emerged as a pivotal strategy for post-training alignment of Multimodal Large Language Models (MLLMs). However, existing approaches predominantly rely on external human-annotated preference data, which is costly and labor-intensive to collect. In this work, we propose OrdMoE, a novel preference alignment framework that bypasses the reliance on external human preferences entirely by leveraging intrinsic signals within Mixture-of-Experts (MoE) architectures. Specifically, we observe that the router's expert selection scores implicitly encode a quality-aware ranking of responses (i.e. higher-scoring experts consistently generate higher-quality outputs). Building on this insight, OrdMoE constructs an internal preference hierarchy by grouping experts into ranked tiers based on their per-token routing scores and activating each tier separately to produce a sequence of responses with increasing quality. This yields a zero-cost, self-supervised preference ordering over generated responses, which can be directly optimized using standard preference learning objectives. Extensive experiments across multiple multimodal benchmarks demnstrate that OrdMoE significantly enhances both alignment and overall performance of multimodal Mixture-of-Experts LLMs, achieving competitive results without requiring any human-annotated preference data.

Yuting Gao, Wang Lan, Hengyuan Zhao et al.

Multimodal Mixture-of-Experts (MoE) models offer a promising path toward scalable and efficient large vision-language systems. However, existing approaches rely on rigid routing strategies (typically activating a fixed number of experts per token) ignoring the inherent heterogeneity in semantic importance across modalities. This leads to suboptimal compute allocation, where redundant tokens consume as many resources as critical ones. To address this, we propose AnyExperts, a novel on-demand, budget-aware dynamic routing framework that allocates a variable total number of expert slots per token based on its semantic importance. Crucially, to prevent uncontrolled compute growth, the total slots per token are constrained within a fixed range, and each slot is filled by either a real expert or a virtual expert, with the virtual share capped at a small maximum (e.g., 20%). The model then adaptively balances the real-to-virtual ratio per token, assigning more real experts to semantically rich regions and relying more on virtual experts for redundant content. Evaluated across diverse tasks in visual understanding, audio understanding, and NLP understanding, AnyExperts improves performance under the same compute budget. Notably, on general image/video tasks, it achieves comparable accuracy with 40% fewer real expert activations; on text-dense tasks (OCR and NLP), it maintains performance while reducing real expert usage by 10%. These results demonstrate that fine-grained, importance-driven expert allocation significantly enhances both the efficiency and effectiveness of multimodal MoE models.

Qixin Sun, Ziqin Wang, Hengyuan Zhao et al.

Retrieval Augmented Generation enhances the response accuracy of Large Language Models (LLMs) by integrating retrieval and generation modules with external knowledge, demonstrating particular strength in real-time queries and Visual Question Answering tasks. However, the effectiveness of RAG is frequently hindered by the precision of the retriever: many retrieved samples fed into the generation phase are irrelevant or misleading, posing a critical bottleneck to LLMs' performance. To address this challenge, we introduce VaccineRAG, a novel Chain-of-Thought-based retrieval-augmented generation dataset. On one hand, VaccineRAG employs a benchmark to evaluate models using data with varying positive/negative sample ratios, systematically exposing inherent weaknesses in current LLMs. On the other hand, it enhances models' sample-discrimination capabilities by prompting LLMs to generate explicit Chain-of-Thought (CoT) analysis for each sample before producing final answers. Furthermore, to enhance the model's ability to learn long-sequence complex CoT content, we propose Partial-GRPO. By modeling the outputs of LLMs as multiple components rather than a single whole, our model can make more informed preference selections for complex sequences, thereby enhancing its capacity to learn complex CoT. Comprehensive evaluations and ablation studies on VaccineRAG validate the effectiveness of the proposed scheme. The code and dataset will be publicly released soon.

Xuzheng Yu, Chen Jiang, Wei Zhang et al.

With the explosive growth of video data in real-world applications, a comprehensive representation of videos becomes increasingly important. In this paper, we address the problem of video scene recognition, whose goal is to learn a high-level video representation to classify scenes in videos. Due to the diversity and complexity of video contents in realistic scenarios, this task remains a challenge. Most existing works identify scenes for videos only from visual or textual information in a temporal perspective, ignoring the valuable information hidden in single frames, while several earlier studies only recognize scenes for separate images in a non-temporal perspective. We argue that these two perspectives are both meaningful for this task and complementary to each other, meanwhile, externally introduced knowledge can also promote the comprehension of videos. We propose a novel two-stream framework to model video representations from multiple perspectives, i.e. temporal and non-temporal perspectives, and integrate the two perspectives in an end-to-end manner by self-distillation. Besides, we design a knowledge-enhanced feature fusion and label prediction method that contributes to naturally introducing knowledge into the task of video scene recognition. Experiments conducted on a real-world dataset demonstrate the effectiveness of our proposed method.