Liqiang Nie

Zixu Li, Yupeng Hu, Zhiwei Chen et al.

Composed Image Retrieval (CIR) represents a challenging retrieval task that targets locating specific images through multimodal inputs. Despite recent progress in CIR techniques, prior approaches often overlook cases where images appear visually alike yet differ in attributes, potentially undermining both multimodal feature fusion and similarity modeling. To mitigate this limitation, we design a unified representation of cross-modal features based on attribute prototypes. Nevertheless, the task is far from straightforward, owing to three core issues: (1) entanglement in attribute-level semantics, (2) inconsistency across modalities, and (3) supervised signal missing. To tackle the above obstacles, we introduce a COMposed image retrieval network guided By attrIbute-based NEighbor Relations (COMBINER). Specifically, we first design an Adaptive Semantic Disentanglement module, which is capable of disentangling attribute features based on multimodal primitive features. Secondly, we propose a Unified Prototype-based Composition module, which can construct cross-modal unified prototypes (CUP) and facilitate multimodal feature composition. Finally, we introduce a Dual Relations Modeling module, which can mine pairwise and neighbor relations based on attribute similarity. Compared to traditional neighbor relations modeling CIR methods, COMBINER represents the first study addressing the phenomenon of visually similar but attribute-unrelated samples. It achieves a more accurate understanding of the semantic relations among samples by employing an attribute prototype-based similarity metric. Comprehensive experiments conducted on three benchmark datasets confirm the effectiveness of our proposed COMBINER. The implementation of our method will be accessed at https://github.com/Lee-zixu/COMBINER

Andrey Moskalenko, Alexey Bryncev, Ivan Kosmynin et al.

This paper presents an overview of the NTIRE 2026 Challenge on Video Saliency Prediction. The goal of the challenge participants was to develop automatic saliency map prediction methods for the provided video sequences. The novel dataset of 2,000 diverse videos with an open license was prepared for this challenge. The fixations and corresponding saliency maps were collected using crowdsourced mouse tracking and contain viewing data from over 5,000 assessors. Evaluation was performed on a subset of 800 test videos using generally accepted quality metrics. The challenge attracted over 20 teams making submissions, and 7 teams passed the final phase with code review. All data used in this challenge is made publicly available - https://github.com/msu-video-group/NTIRE26_Saliency_Prediction.

Rui Shao, Tianxing Wu, Liqiang Nie et al.

Existing deepfake detection methods fail to generalize well to unseen or degraded samples, which can be attributed to the over-fitting of low-level forgery patterns. Here we argue that high-level semantics are also indispensable recipes for generalizable forgery detection. Recently, large pre-trained Vision Transformers (ViTs) have shown promising generalization capability. In this paper, we propose the first parameter-efficient tuning approach for deepfake detection, namely DeepFake-Adapter, to effectively and efficiently adapt the generalizable high-level semantics from large pre-trained ViTs to aid deepfake detection. Given large pre-trained models but limited deepfake data, DeepFake-Adapter introduces lightweight yet dedicated dual-level adapter modules to a ViT while keeping the model backbone frozen. Specifically, to guide the adaptation process to be aware of both global and local forgery cues of deepfake data, 1) we not only insert Globally-aware Bottleneck Adapters in parallel to MLP layers of ViT, 2) but also actively cross-attend Locally-aware Spatial Adapters with features from ViT. Unlike existing deepfake detection methods merely focusing on low-level forgery patterns, the forgery detection process of our model can be regularized by generalizable high-level semantics from a pre-trained ViT and adapted by global and local low-level forgeries of deepfake data. Extensive experiments on several standard deepfake detection benchmarks validate the effectiveness of our approach. Notably, DeepFake-Adapter demonstrates a convincing advantage under cross-dataset and cross-manipulation settings. The code has been released at https://github.com/rshaojimmy/DeepFake-Adapter.

Teng Sun, Juntong Ni, Wenjie Wang et al.

Existing work on Multimodal Sentiment Analysis (MSA) utilizes multimodal information for prediction yet unavoidably suffers from fitting the spurious correlations between multimodal features and sentiment labels. For example, if most videos with a blue background have positive labels in a dataset, the model will rely on such correlations for prediction, while "blue background" is not a sentiment-related feature. To address this problem, we define a general debiasing MSA task, which aims to enhance the Out-Of-Distribution (OOD) generalization ability of MSA models by reducing their reliance on spurious correlations. To this end, we propose a general debiasing framework based on Inverse Probability Weighting (IPW), which adaptively assigns small weights to the samples with larger bias (i.e., the severer spurious correlations). The key to this debiasing framework is to estimate the bias of each sample, which is achieved by two steps: 1) disentangling the robust features and biased features in each modality, and 2) utilizing the biased features to estimate the bias. Finally, we employ IPW to reduce the effects of large-biased samples, facilitating robust feature learning for sentiment prediction. To examine the model's generalization ability, we keep the original testing sets on two benchmarks and additionally construct multiple unimodal and multimodal OOD testing sets. The empirical results demonstrate the superior generalization ability of our proposed framework. We have released the code and data to facilitate the reproduction https://github.com/Teng-Sun/GEAR.

Yali Du, Yinwei Wei, Wei Ji et al.

The booming development and huge market of micro-videos bring new e-commerce channels for merchants. Currently, more micro-video publishers prefer to embed relevant ads into their micro-videos, which not only provides them with business income but helps the audiences to discover their interesting products. However, due to the micro-video recording by unprofessional equipment, involving various topics and including multiple modalities, it is challenging to locate the products related to micro-videos efficiently, appropriately, and accurately. We formulate the microvideo-product retrieval task, which is the first attempt to explore the retrieval between the multi-modal and multi-modal instances. A novel approach named Multi-Queue Momentum Contrast (MQMC) network is proposed for bidirectional retrieval, consisting of the uni-modal feature and multi-modal instance representation learning. Moreover, a discriminative selection strategy with a multi-queue is used to distinguish the importance of different negatives based on their categories. We collect two large-scale microvideo-product datasets (MVS and MVS-large) for evaluation and manually construct the hierarchical category ontology, which covers sundry products in daily life. Extensive experiments show that MQMC outperforms the state-of-the-art baselines. Our replication package (including code, dataset, etc.) is publicly available at https://github.com/duyali2000/MQMC.

Yangyang Guo, Fangkai Jiao, Zhiqi Shen et al.

Teaching Visual Question Answering (VQA) models to refrain from answering unanswerable questions is necessary for building a trustworthy AI system. Existing studies, though have explored various aspects of VQA but somewhat ignored this particular attribute. This paper aims to bridge the research gap by contributing a comprehensive dataset, called UNK-VQA. The dataset is specifically designed to address the challenge of questions that models do not know. To this end, we first augment the existing data via deliberate perturbations on either the image or question. In specific, we carefully ensure that the question-image semantics remain close to the original unperturbed distribution. By this means, the identification of unanswerable questions becomes challenging, setting our dataset apart from others that involve mere image replacement. We then extensively evaluate the zero- and few-shot performance of several emerging multi-modal large models and discover their significant limitations when applied to our dataset. Additionally, we also propose a straightforward method to tackle these unanswerable questions. This dataset, we believe, will serve as a valuable benchmark for enhancing the abstention capability of VQA models, thereby leading to increased trustworthiness of AI systems. We have made the dataset (https://github.com/guoyang9/UNK-VQA) available to facilitate further exploration in this area.

Zixu Li, Yupeng Hu, Zhiheng Fu et al.

The CoVR-R challenge evaluates composed video retrieval, where a system must retrieve a target video from a large gallery given a reference video and a textual edit instruction. This setting is not a standard video-text retrieval problem: the query is defined by both the visual evidence in the source video and the transformation implied by the edit. A strong embedding model can provide scalable candidate recall, but it may under-express target-side consequences such as state changes, action replacement, object preservation, or temporal consistency. A pairwise multimodal reranker can verify such details more directly, but exhaustive reranking over the full gallery is computationally infeasible. We present $\mathbb{R}^3$, a zero-shot composed video retrieval pipeline built around Reasoning-guided Recalling and Reranking. The core idea is to turn the source-edit query into a reasoning-grounded retrieval program rather than treating the edit text as a short caption. First, the model generates a reasoning trace that describes the expected target video after applying the edit. Then the trace is encoded together with the source video as a reasoning-augmented query, and its retrieval score is fused with the base composed query through an agreement-gated residual rule. At last, a re-ranker verifies the recalled candidates with direct source-candidate comparison. Experiments have demonstrated the effectiveness of our method in addressing this challenge. Codes are available on https://github.com/Lee-zixu/R-3.

Peiguang Jing, Xianyi Liu, Ji Wang et al.

Emotion distribution learning has gained increasing attention with the tendency to express emotions through images. As for emotion ambiguity arising from humans' subjectivity, substantial previous methods generally focused on learning appropriate representations from the holistic or significant part of images. However, they rarely consider establishing connections with the stylistic information although it can lead to a better understanding of images. In this paper, we propose a style-guided high-order attention network for image emotion distribution learning termed StyleEDL, which interactively learns stylistic-aware representations of images by exploring the hierarchical stylistic information of visual contents. Specifically, we consider exploring the intra- and inter-layer correlations among GRAM-based stylistic representations, and meanwhile exploit an adversary-constrained high-order attention mechanism to capture potential interactions between subtle visual parts. In addition, we introduce a stylistic graph convolutional network to dynamically generate the content-dependent emotion representations to benefit the final emotion distribution learning. Extensive experiments conducted on several benchmark datasets demonstrate the effectiveness of our proposed StyleEDL compared to state-of-the-art methods. The implementation is released at: https://github.com/liuxianyi/StyleEDL.

Linmei Hu, Zeyi Liu, Ziwang Zhao et al.

Pre-trained Language Models (PLMs) which are trained on large text corpus via self-supervised learning method, have yielded promising performance on various tasks in Natural Language Processing (NLP). However, though PLMs with huge parameters can effectively possess rich knowledge learned from massive training text and benefit downstream tasks at the fine-tuning stage, they still have some limitations such as poor reasoning ability due to the lack of external knowledge. Research has been dedicated to incorporating knowledge into PLMs to tackle these issues. In this paper, we present a comprehensive review of Knowledge Enhanced Pre-trained Language Models (KE-PLMs) to provide a clear insight into this thriving field. We introduce appropriate taxonomies respectively for Natural Language Understanding (NLU) and Natural Language Generation (NLG) to highlight these two main tasks of NLP. For NLU, we divide the types of knowledge into four categories: linguistic knowledge, text knowledge, knowledge graph (KG), and rule knowledge. The KE-PLMs for NLG are categorized into KG-based and retrieval-based methods. Finally, we point out some promising future directions of KE-PLMs.

Leigang Qu, Shengqiong Wu, Hao Fei et al.

In the text-to-image generation field, recent remarkable progress in Stable Diffusion makes it possible to generate rich kinds of novel photorealistic images. However, current models still face misalignment issues (e.g., problematic spatial relation understanding and numeration failure) in complex natural scenes, which impedes the high-faithfulness text-to-image generation. Although recent efforts have been made to improve controllability by giving fine-grained guidance (e.g., sketch and scribbles), this issue has not been fundamentally tackled since users have to provide such guidance information manually. In this work, we strive to synthesize high-fidelity images that are semantically aligned with a given textual prompt without any guidance. Toward this end, we propose a coarse-to-fine paradigm to achieve layout planning and image generation. Concretely, we first generate the coarse-grained layout conditioned on a given textual prompt via in-context learning based on Large Language Models. Afterward, we propose a fine-grained object-interaction diffusion method to synthesize high-faithfulness images conditioned on the prompt and the automatically generated layout. Extensive experiments demonstrate that our proposed method outperforms the state-of-the-art models in terms of layout and image generation. Our code and settings are available at https://layoutllm-t2i.github.io.

Xingning Dong, Tian Gan, Xuemeng Song et al.

Scene Graph Generation, which generally follows a regular encoder-decoder pipeline, aims to first encode the visual contents within the given image and then parse them into a compact summary graph. Existing SGG approaches generally not only neglect the insufficient modality fusion between vision and language, but also fail to provide informative predicates due to the biased relationship predictions, leading SGG far from practical. Towards this end, in this paper, we first present a novel Stacked Hybrid-Attention network, which facilitates the intra-modal refinement as well as the inter-modal interaction, to serve as the encoder. We then devise an innovative Group Collaborative Learning strategy to optimize the decoder. Particularly, based upon the observation that the recognition capability of one classifier is limited towards an extremely unbalanced dataset, we first deploy a group of classifiers that are expert in distinguishing different subsets of classes, and then cooperatively optimize them from two aspects to promote the unbiased SGG. Experiments conducted on VG and GQA datasets demonstrate that, we not only establish a new state-of-the-art in the unbiased metric, but also nearly double the performance compared with two baselines.

Min Cao, Shiping Li, Juntao Li et al.

In the past few years, cross-modal image-text retrieval (ITR) has experienced increased interest in the research community due to its excellent research value and broad real-world application. It is designed for the scenarios where the queries are from one modality and the retrieval galleries from another modality. This paper presents a comprehensive and up-to-date survey on the ITR approaches from four perspectives. By dissecting an ITR system into two processes: feature extraction and feature alignment, we summarize the recent advance of the ITR approaches from these two perspectives. On top of this, the efficiency-focused study on the ITR system is introduced as the third perspective. To keep pace with the times, we also provide a pioneering overview of the cross-modal pre-training ITR approaches as the fourth perspective. Finally, we outline the common benchmark datasets and valuation metric for ITR, and conduct the accuracy comparison among the representative ITR approaches. Some critical yet less studied issues are discussed at the end of the paper.

Tianyi Wang, Harry Cheng, Kam Pui Chow et al.

Recently, Deepfake has drawn considerable public attention due to security and privacy concerns in social media digital forensics. As the wildly spreading Deepfake videos on the Internet become more realistic, traditional detection techniques have failed in distinguishing between real and fake. Most existing deep learning methods mainly focus on local features and relations within the face image using convolutional neural networks as a backbone. However, local features and relations are insufficient for model training to learn enough general information for Deepfake detection. Therefore, the existing Deepfake detection methods have reached a bottleneck to further improve the detection performance. To address this issue, we propose a deep convolutional Transformer to incorporate the decisive image features both locally and globally. Specifically, we apply convolutional pooling and re-attention to enrich the extracted features and enhance efficacy. Moreover, we employ the barely discussed image keyframes in model training for performance improvement and visualize the feature quantity gap between the key and normal image frames caused by video compression. We finally illustrate the transferability with extensive experiments on several Deepfake benchmark datasets. The proposed solution consistently outperforms several state-of-the-art baselines on both within- and cross-dataset experiments.

Harry Cheng, Yangyang Guo, Tianyi Wang et al.

Detecting forgery videos is highly desirable due to the abuse of deepfake. Existing detection approaches contribute to exploring the specific artifacts in deepfake videos and fit well on certain data. However, the growing technique on these artifacts keeps challenging the robustness of traditional deepfake detectors. As a result, the development of generalizability of these approaches has reached a blockage. To address this issue, given the empirical results that the identities behind voices and faces are often mismatched in deepfake videos, and the voices and faces have homogeneity to some extent, in this paper, we propose to perform the deepfake detection from an unexplored voice-face matching view. To this end, a voice-face matching method is devised to measure the matching degree of these two. Nevertheless, training on specific deepfake datasets makes the model overfit certain traits of deepfake algorithms. We instead, advocate a method that quickly adapts to untapped forgery, with a pre-training then fine-tuning paradigm. Specifically, we first pre-train the model on a generic audio-visual dataset, followed by the fine-tuning on downstream deepfake data. We conduct extensive experiments over three widely exploited deepfake datasets - DFDC, FakeAVCeleb, and DeepfakeTIMIT. Our method obtains significant performance gains as compared to other state-of-the-art competitors. It is also worth noting that our method already achieves competitive results when fine-tuned on limited deepfake data.

Chengliang Liu, Jie Wen, Yong Xu et al.

As a cross-topic of multi-view learning and multi-label classification, multi-view multi-label classification has gradually gained traction in recent years. The application of multi-view contrastive learning has further facilitated this process, however, the existing multi-view contrastive learning methods crudely separate the so-called negative pair, which largely results in the separation of samples belonging to the same category or similar ones. Besides, plenty of multi-view multi-label learning methods ignore the possible absence of views and labels. To address these issues, in this paper, we propose an incomplete multi-view missing multi-label classification network named RANK. In this network, a label-driven multi-view contrastive learning strategy is proposed to leverage supervised information to preserve the intra-view structure and perform the cross-view consistency alignment. Furthermore, we break through the view-level weights inherent in existing methods and propose a quality-aware sub-network to dynamically assign quality scores to each view of each sample. The label correlation information is fully utilized in the final multi-label cross-entropy classification loss, effectively improving the discriminative power. Last but not least, our model is not only able to handle complete multi-view multi-label data, but also works on datasets with missing instances and labels. Extensive experiments confirm that our RANK outperforms existing state-of-the-art methods.

Leyang Shen, Gongwei Chen, Rui Shao et al.

Multimodal large language models (MLLMs) have demonstrated impressive capabilities across various vision-language tasks. However, a generalist MLLM typically underperforms compared with a specialist MLLM on most VL tasks, which can be attributed to task interference. In this paper, we propose a mixture of multimodal experts (MoME) to mitigate task interference and obtain a generalist MLLM. Our MoME is composed of two key components, a mixture of vision experts (MoVE) and a mixture of language experts (MoLE). MoVE can adaptively modulate the features transformed from various vision encoders, and has a strong compatibility in transformation architecture. MoLE incorporates sparsely gated experts into LLMs to achieve painless improvements with roughly unchanged inference costs. In response to task interference, our MoME specializes in both vision and language modality to adapt to task discrepancies. Extensive experiments show that MoME significantly improves the performance of generalist MLLMs across various VL tasks. The source code is released at https://github.com/JiuTian-VL/MoME

Zhonghua Zheng, Lizi Liao, Yang Deng et al.

The integration of emotional support into various conversational scenarios presents profound societal benefits, such as social interactions, mental health counseling, and customer service. However, there are unsolved challenges that hinder real-world applications in this field, including limited data availability and the absence of well-accepted model training paradigms. This work endeavors to navigate these challenges by harnessing the capabilities of Large Language Models (LLMs). We introduce an innovative methodology that synthesizes human insights with the computational prowess of LLMs to curate an extensive emotional support dialogue dataset. Our approach is initiated with a meticulously designed set of dialogues spanning diverse scenarios as generative seeds. By utilizing the in-context learning potential of ChatGPT, we recursively generate an ExTensible Emotional Support dialogue dataset, named ExTES. Following this, we deploy advanced tuning techniques on the LLaMA model, examining the impact of diverse training strategies, ultimately yielding an LLM meticulously optimized for emotional support interactions. An exhaustive assessment of the resultant model showcases its proficiency in offering emotional support, marking a pivotal step in the realm of emotional support bots and paving the way for subsequent research and implementations.

Rui Shao, Tianxing Wu, Jianlong Wu et al.

Misinformation has become a pressing issue. Fake media, in both visual and textual forms, is widespread on the web. While various deepfake detection and text fake news detection methods have been proposed, they are only designed for single-modality forgery based on binary classification, let alone analyzing and reasoning subtle forgery traces across different modalities. In this paper, we highlight a new research problem for multi-modal fake media, namely Detecting and Grounding Multi-Modal Media Manipulation (DGM^4). DGM^4 aims to not only detect the authenticity of multi-modal media, but also ground the manipulated content, which requires deeper reasoning of multi-modal media manipulation. To support a large-scale investigation, we construct the first DGM^4 dataset, where image-text pairs are manipulated by various approaches, with rich annotation of diverse manipulations. Moreover, we propose a novel HierArchical Multi-modal Manipulation rEasoning tRansformer (HAMMER) to fully capture the fine-grained interaction between different modalities. HAMMER performs 1) manipulation-aware contrastive learning between two uni-modal encoders as shallow manipulation reasoning, and 2) modality-aware cross-attention by multi-modal aggregator as deep manipulation reasoning. Dedicated manipulation detection and grounding heads are integrated from shallow to deep levels based on the interacted multi-modal information. To exploit more fine-grained contrastive learning for cross-modal semantic alignment, we further integrate Manipulation-Aware Contrastive Loss with Local View and construct a more advanced model HAMMER++. Finally, we build an extensive benchmark and set up rigorous evaluation metrics for this new research problem. Comprehensive experiments demonstrate the superiority of HAMMER and HAMMER++.

Teng Sun, Wenjie Wang, Liqiang Jing et al.

Existing studies on multimodal sentiment analysis heavily rely on textual modality and unavoidably induce the spurious correlations between textual words and sentiment labels. This greatly hinders the model generalization ability. To address this problem, we define the task of out-of-distribution (OOD) multimodal sentiment analysis. This task aims to estimate and mitigate the bad effect of textual modality for strong OOD generalization. To this end, we embrace causal inference, which inspects the causal relationships via a causal graph. From the graph, we find that the spurious correlations are attributed to the direct effect of textual modality on the model prediction while the indirect one is more reliable by considering multimodal semantics. Inspired by this, we devise a model-agnostic counterfactual framework for multimodal sentiment analysis, which captures the direct effect of textual modality via an extra text model and estimates the indirect one by a multimodal model. During the inference, we first estimate the direct effect by the counterfactual inference, and then subtract it from the total effect of all modalities to obtain the indirect effect for reliable prediction. Extensive experiments show the superior effectiveness and generalization ability of our proposed framework.

Yangyang Guo, Liqiang Nie, Yongkang Wong et al.

Knowledge-based Visual Question Answering (VQA) expects models to rely on external knowledge for robust answer prediction. Though significant it is, this paper discovers several leading factors impeding the advancement of current state-of-the-art methods. On the one hand, methods which exploit the explicit knowledge take the knowledge as a complement for the coarsely trained VQA model. Despite their effectiveness, these approaches often suffer from noise incorporation and error propagation. On the other hand, pertaining to the implicit knowledge, the multi-modal implicit knowledge for knowledge-based VQA still remains largely unexplored. This work presents a unified end-to-end retriever-reader framework towards knowledge-based VQA. In particular, we shed light on the multi-modal implicit knowledge from vision-language pre-training models to mine its potential in knowledge reasoning. As for the noise problem encountered by the retrieval operation on explicit knowledge, we design a novel scheme to create pseudo labels for effective knowledge supervision. This scheme is able to not only provide guidance for knowledge retrieval, but also drop these instances potentially error-prone towards question answering. To validate the effectiveness of the proposed method, we conduct extensive experiments on the benchmark dataset. The experimental results reveal that our method outperforms existing baselines by a noticeable margin. Beyond the reported numbers, this paper further spawns several insights on knowledge utilization for future research with some empirical findings.

Rui Hao, Linmei Hu, Weijian Qi et al.

Dialogue-based language models mark a huge milestone in the field of artificial intelligence, by their impressive ability to interact with users, as well as a series of challenging tasks prompted by customized instructions. However, the prevalent large-scale dialogue-based language models like ChatGPT still have room for improvement, such as unstable responses to questions and the inability to think cooperatively like humans. Considering the ability of dialogue-based language models in conversation and their inherent randomness in thinking, we propose ChatLLM network that allows multiple dialogue-based language models to interact, provide feedback, and think together. We design the network of ChatLLMs based on ChatGPT. Specifically, individual instances of ChatGPT may possess distinct perspectives towards the same problem, and by consolidating these diverse viewpoints via a separate ChatGPT, the ChatLLM network system can conduct decision-making more objectively and comprehensively. In addition, a language-based feedback mechanism comparable to backpropagation is devised to update the ChatGPTs within the network. Experiments on two datasets demonstrate that our network attains significant improvements in problem-solving, leading to observable progress amongst each member.

Fangkai Jiao, Yangyang Guo, Xuemeng Song et al.

Logical reasoning is of vital importance to natural language understanding. Previous studies either employ graph-based models to incorporate prior knowledge about logical relations, or introduce symbolic logic into neural models through data augmentation. These methods, however, heavily depend on annotated training data, and thus suffer from over-fitting and poor generalization problems due to the dataset sparsity. To address these two problems, in this paper, we propose MERIt, a MEta-path guided contrastive learning method for logical ReasonIng of text, to perform self-supervised pre-training on abundant unlabeled text data. Two novel strategies serve as indispensable components of our method. In particular, a strategy based on meta-path is devised to discover the logical structure in natural texts, followed by a counterfactual data augmentation strategy to eliminate the information shortcut induced by pre-training. The experimental results on two challenging logical reasoning benchmarks, i.e., ReClor and LogiQA, demonstrate that our method outperforms the SOTA baselines with significant improvements.

Leigang Qu, Meng Liu, Wenjie Wang et al.

Image-text retrieval aims to bridge the modality gap and retrieve cross-modal content based on semantic similarities. Prior work usually focuses on the pairwise relations (i.e., whether a data sample matches another) but ignores the higher-order neighbor relations (i.e., a matching structure among multiple data samples). Re-ranking, a popular post-processing practice, has revealed the superiority of capturing neighbor relations in single-modality retrieval tasks. However, it is ineffective to directly extend existing re-ranking algorithms to image-text retrieval. In this paper, we analyze the reason from four perspectives, i.e., generalization, flexibility, sparsity, and asymmetry, and propose a novel learnable pillar-based re-ranking paradigm. Concretely, we first select top-ranked intra- and inter-modal neighbors as pillars, and then reconstruct data samples with the neighbor relations between them and the pillars. In this way, each sample can be mapped into a multimodal pillar space only using similarities, ensuring generalization. After that, we design a neighbor-aware graph reasoning module to flexibly exploit the relations and excavate the sparse positive items within a neighborhood. We also present a structure alignment constraint to promote cross-modal collaboration and align the asymmetric modalities. On top of various base backbones, we carry out extensive experiments on two benchmark datasets, i.e., Flickr30K and MS-COCO, demonstrating the effectiveness, superiority, generalization, and transferability of our proposed re-ranking paradigm.

Meng Liu, Liqiang Nie, Yunxiao Wang et al.

Video moment localization, also known as video moment retrieval, aiming to search a target segment within a video described by a given natural language query. Beyond the task of temporal action localization whereby the target actions are pre-defined, video moment retrieval can query arbitrary complex activities. In this survey paper, we aim to present a comprehensive review of existing video moment localization techniques, including supervised, weakly supervised, and unsupervised ones. We also review the datasets available for video moment localization and group results of related work. In addition, we discuss promising future directions for this field, in particular large-scale datasets and interpretable video moment localization models.

Xiaolin Chen, Xuemeng Song, Liqiang Jing et al.

Text response generation for multimodal task-oriented dialog systems, which aims to generate the proper text response given the multimodal context, is an essential yet challenging task. Although existing efforts have achieved compelling success, they still suffer from two pivotal limitations: 1) overlook the benefit of generative pre-training, and 2) ignore the textual context related knowledge. To address these limitations, we propose a novel dual knowledge-enhanced generative pretrained language model for multimodal task-oriented dialog systems (DKMD), consisting of three key components: dual knowledge selection, dual knowledge-enhanced context learning, and knowledge-enhanced response generation. To be specific, the dual knowledge selection component aims to select the related knowledge according to both textual and visual modalities of the given context. Thereafter, the dual knowledge-enhanced context learning component targets seamlessly integrating the selected knowledge into the multimodal context learning from both global and local perspectives, where the cross-modal semantic relation is also explored. Moreover, the knowledge-enhanced response generation component comprises a revised BART decoder, where an additional dot-product knowledge-decoder attention sub-layer is introduced for explicitly utilizing the knowledge to advance the text response generation. Extensive experiments on a public dataset verify the superiority of the proposed DKMD over state-of-the-art competitors.

Liqiang Jing, Xuemeng Song, Kun Ouyang et al.

Multimodal Sarcasm Explanation (MuSE) is a new yet challenging task, which aims to generate a natural language sentence for a multimodal social post (an image as well as its caption) to explain why it contains sarcasm. Although the existing pioneer study has achieved great success with the BART backbone, it overlooks the gap between the visual feature space and the decoder semantic space, the object-level metadata of the image, as well as the potential external knowledge. To solve these limitations, in this work, we propose a novel mulTi-source sEmantic grAph-based Multimodal sarcasm explanation scheme, named TEAM. In particular, TEAM extracts the object-level semantic meta-data instead of the traditional global visual features from the input image. Meanwhile, TEAM resorts to ConceptNet to obtain the external related knowledge concepts for the input text and the extracted object meta-data. Thereafter, TEAM introduces a multi-source semantic graph that comprehensively characterize the multi-source (i.e., caption, object meta-data, external knowledge) semantic relations to facilitate the sarcasm reasoning. Extensive experiments on a public released dataset MORE verify the superiority of our model over cutting-edge methods.

Yuzhang Shang, Dan Xu, Bin Duan et al.

Relying on the premise that the performance of a binary neural network can be largely restored with eliminated quantization error between full-precision weight vectors and their corresponding binary vectors, existing works of network binarization frequently adopt the idea of model robustness to reach the aforementioned objective. However, robustness remains to be an ill-defined concept without solid theoretical support. In this work, we introduce the Lipschitz continuity, a well-defined functional property, as the rigorous criteria to define the model robustness for BNN. We then propose to retain the Lipschitz continuity as a regularization term to improve the model robustness. Particularly, while the popular Lipschitz-involved regularization methods often collapse in BNN due to its extreme sparsity, we design the Retention Matrices to approximate spectral norms of the targeted weight matrices, which can be deployed as the approximation for the Lipschitz constant of BNNs without the exact Lipschitz constant computation (NP-hard). Our experiments prove that our BNN-specific regularization method can effectively strengthen the robustness of BNN (testified on ImageNet-C), achieving state-of-the-art performance on CIFAR and ImageNet.

Yuzhang Shang, Dan Xu, Ziliang Zong et al.

Neural network binarization accelerates deep models by quantizing their weights and activations into 1-bit. However, there is still a huge performance gap between Binary Neural Networks (BNNs) and their full-precision (FP) counterparts. As the quantization error caused by weights binarization has been reduced in earlier works, the activations binarization becomes the major obstacle for further improvement of the accuracy. BNN characterises a unique and interesting structure, where the binary and latent FP activations exist in the same forward pass (i.e., $\text{Binarize}(\mathbf{a}_F) = \mathbf{a}_B$). To mitigate the information degradation caused by the binarization operation from FP to binary activations, we establish a novel contrastive learning framework while training BNNs through the lens of Mutual Information (MI) maximization. MI is introduced as the metric to measure the information shared between binary and FP activations, which assists binarization with contrastive learning. Specifically, the representation ability of the BNNs is greatly strengthened via pulling the positive pairs with binary and FP activations from the same input samples, as well as pushing negative pairs from different samples (the number of negative pairs can be exponentially large). This benefits the downstream tasks, not only classification but also segmentation and depth estimation, etc. The experimental results show that our method can be implemented as a pile-up module on existing state-of-the-art binarization methods and can remarkably improve the performance over them on CIFAR-10/100 and ImageNet, in addition to the great generalization ability on NYUD-v2.

Zixu Li, Zhiwei Chen, Zhiheng Fu et al.

The 1st Cross-Domain EgoCross Challenge at EgoVis, CVPR 2026 evaluates whether multimodal large language models can reason over egocentric videos across surgery, industry, extreme sports, and animal perspective. We achieved second place in both the Source-Limited and Open-Source tracks. In this report, we formulate EgoCross as a robust cross-domain embodied video reasoning problem rather than a simple multiple-choice visual question answering task. We identify three key challenges: (C1) temporal boundary ambiguity, where critical state transitions are sparsely sampled and often occur between frames; (C2) cross-domain semantic granularity mismatch, where the same capability requires different domain-specific visual grammar; and (C3) decision instability under close options, where long multimodal reasoning can select unsupported distractors or produce malformed outputs. To address them, we propose OmniEgo-R$^2$ (Omnidomain Egocentric Routed Reasoning), a unified routed reasoning pipeline consisting of temporal-evidence normalization, domain-agnostic capability routing, structured perception--dynamics--decision reasoning, boundary-aware option verification, and defensive answer calibration. OmniEgo-R$^2$ uses the Qwen3-VL-4B-SFT checkpoints on each EgoCross domain as the visual-language backbone, and wraps them with lightweight test-time reasoning and parsing programs. Our final submissions obtain 66.35% overall accuracy in the Source-Limited track and 66.77% in the Open-Source track, ranking second in both leaderboards.

Yudong Han, Liqiang Nie, Jianhua Yin et al.

Several studies have recently pointed that existing Visual Question Answering (VQA) models heavily suffer from the language prior problem, which refers to capturing superficial statistical correlations between the question type and the answer whereas ignoring the image contents. Numerous efforts have been dedicated to strengthen the image dependency by creating the delicate models or introducing the extra visual annotations. However, these methods cannot sufficiently explore how the visual cues explicitly affect the learned answer representation, which is vital for language reliance alleviation. Moreover, they generally emphasize the class-level discrimination of the learned answer representation, which overlooks the more fine-grained instance-level patterns and demands further optimization. In this paper, we propose a novel collaborative learning scheme from the viewpoint of visual perturbation calibration, which can better investigate the fine-grained visual effects and mitigate the language prior problem by learning the instance-level characteristics. Specifically, we devise a visual controller to construct two sorts of curated images with different perturbation extents, based on which the collaborative learning of intra-instance invariance and inter-instance discrimination is implemented by two well-designed discriminators. Besides, we implement the information bottleneck modulator on latent space for further bias alleviation and representation calibration. We impose our visual perturbation-aware framework to three orthodox baselines and the experimental results on two diagnostic VQA-CP benchmark datasets evidently demonstrate its effectiveness. In addition, we also justify its robustness on the balanced VQA benchmark.

Zhenyang Li, Yangyang Guo, Kejie Wang et al.

Visual Commonsense Reasoning (VCR) remains a significant yet challenging research problem in the realm of visual reasoning. A VCR model generally aims at answering a textual question regarding an image, followed by the rationale prediction for the preceding answering process. Though these two processes are sequential and intertwined, existing methods always consider them as two independent matching-based instances. They, therefore, ignore the pivotal relationship between the two processes, leading to sub-optimal model performance. This paper presents a novel visual attention alignment method to efficaciously handle these two processes in a unified framework. To achieve this, we first design a re-attention module for aggregating the vision attention map produced in each process. Thereafter, the resultant two sets of attention maps are carefully aligned to guide the two processes to make decisions based on the same image regions. We apply this method to both conventional attention and the recent Transformer models and carry out extensive experiments on the VCR benchmark dataset. The results demonstrate that with the attention alignment module, our method achieves a considerable improvement over the baseline methods, evidently revealing the feasibility of the coupling of the two processes as well as the effectiveness of the proposed method.

Yudong Han, Jianhua Yin, Jianlong Wu et al.

Visual Question Answering (VQA) is fundamentally compositional in nature, and many questions are simply answered by decomposing them into modular sub-problems. The recent proposed Neural Module Network (NMN) employ this strategy to question answering, whereas heavily rest with off-the-shelf layout parser or additional expert policy regarding the network architecture design instead of learning from the data. These strategies result in the unsatisfactory adaptability to the semantically-complicated variance of the inputs, thereby hindering the representational capacity and generalizability of the model. To tackle this problem, we propose a Semantic-aware modUlar caPsulE Routing framework, termed as SUPER, to better capture the instance-specific vision-semantic characteristics and refine the discriminative representations for prediction. Particularly, five powerful specialized modules as well as dynamic routers are tailored in each layer of the SUPER network, and the compact routing spaces are constructed such that a variety of customizable routes can be sufficiently exploited and the vision-semantic representations can be explicitly calibrated. We comparatively justify the effectiveness and generalization ability of our proposed SUPER scheme over five benchmark datasets, as well as the parametric-efficient advantage. It is worth emphasizing that this work is not to pursue the state-of-the-art results in VQA. Instead, we expect that our model is responsible to provide a novel perspective towards architecture learning and representation calibration for VQA.

Xiao Wang, Tian Gan, Yinwei Wei et al.

The last decade has witnessed the proliferation of micro-videos on various user-generated content platforms. According to our statistics, around 85.7\% of micro-videos lack annotation. In this paper, we focus on annotating micro-videos with tags. Existing methods mostly focus on analyzing video content, neglecting users' social influence and tag relation. Meanwhile, existing tag relation construction methods suffer from either deficient performance or low tag coverage. To jointly model social influence and tag relation, we formulate micro-video tagging as a link prediction problem in a constructed heterogeneous network. Specifically, the tag relation (represented by tag ontology) is constructed in a semi-supervised manner. Then, we combine tag relation, video-tag annotation, and user-follow relation to build the network. Afterward, a better video and tag representation are derived through Behavior Spread modeling and visual and linguistic knowledge aggregation. Finally, the semantic similarity between each micro-video and all candidate tags is calculated in this video-tag network. Extensive experiments on industrial datasets of three verticals verify the superiority of our model compared with several state-of-the-art baselines.

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.

Harry Cheng, Yangyang Guo, Liqiang Nie et al.

Training an effective video action recognition model poses significant computational challenges, particularly under limited resource budgets. Current methods primarily aim to either reduce model size or utilize pre-trained models, limiting their adaptability to various backbone architectures. This paper investigates the issue of over-sampled frames, a prevalent problem in many approaches yet it has received relatively little attention. Despite the use of fewer frames being a potential solution, this approach often results in a substantial decline in performance. To address this issue, we propose a novel method to restore the intermediate features for two sparsely sampled and adjacent video frames. This feature restoration technique brings a negligible increase in computational requirements compared to resource-intensive image encoders, such as ViT. To evaluate the effectiveness of our method, we conduct extensive experiments on four public datasets, including Kinetics-400, ActivityNet, UCF-101, and HMDB-51. With the integration of our method, the efficiency of three commonly used baselines has been improved by over 50%, with a mere 0.5% reduction in recognition accuracy. In addition, our method also surprisingly helps improve the generalization ability of the models under zero-shot settings.

Xiaojie Li, Yibo Yang, Jianlong Wu et al.

Few-shot class-incremental learning (FSCIL) aims to incrementally learn novel classes from limited examples while preserving knowledge of previously learned classes. Existing methods face a critical dilemma: static architectures rely on a fixed parameter space to learn from data that arrive sequentially, prone to overfitting to the current session, while dynamic architectures require the expansion of the parameter space continually, leading to increased complexity. In this study, we explore the potential of Selective State Space Models (SSMs) for FSCIL. Mamba leverages its input-dependent parameters to dynamically adjust its processing patterns and generate content-aware scan patterns within a fixed architecture. This enables it to configure distinct processing for base and novel classes, effectively preserving existing knowledge while adapting to new ones. To leverage Mamba's potential for FSCIL, we design two key modules: First, we propose a dual selective SSM projector that dynamically adjusts the projection parameters based on the intermediate features for dynamic adaptation. The dual-design structurally decouples base and novel class processing with a frozen base branch, employing a frozen base branch to maintain robust base-class features and a dynamic incremental branch that adaptively learns distinctive feature shifts for novel classes. Second, we develop a class-sensitive selective scan mechanism to guide dynamic adaptation of the incremental branch. It minimizes the disruption to base-class representations caused by training on novel data, and meanwhile, forces the selective scan to perform in distinct patterns between base and novel classes. Extensive experiments on miniImageNet, CUB-200, and CIFAR-100 demonstrate that Mamba-FSCIL achieves state-of-the-art performance. The code is available at https://github.com/xiaojieli0903/Mamba-FSCIL.

Qiaohui Chu, Haoyu Zhang, Yisen Feng et al.

We propose JFAA, a JEPA-based Future Action Anticipation method for the EPIC-KITCHENS-100 (EK-100) Action Anticipation task. Inspired by the representation learning and future prediction ability of V-JEPA 2.1, JFAA uses a frozen encoder and predictor to extract observed context features and near-future latent tokens. A lightweight attentive probe is then trained to predict verb, noun, and action logits with separate task queries. To improve robustness, we further build a field-aware ensemble over selected epoch-level predictions, allowing each output field to benefit from its most reliable candidates. Experimental results on the official challenge server show that JFAA achieves first place in the EgoVis 2026 EK-100 Action Anticipation Challenge. Our code will be released at https://github.com/CorrineQiu/JFAA.

Qiaohui Chu, Haoyu Zhang, Yisen Feng et al.

We propose VISTA, a V-JEPA Integrated StillFast Temporal Anticipator for the Ego4D Short-Term Object Interaction Anticipation (STA) Challenge at EgoVis 2026. Given an egocentric video timestamp, the task requires anticipating the next human-object interaction, including the future active object's bounding box, noun category, verb category, time-to-contact, and confidence score. VISTA follows a StillFast-style design that combines object-centric spatial detection with short-horizon temporal context. Specifically, a COCO-pretrained Faster R-CNN ResNet-50 FPN detector generates object proposals from the last observed high-resolution frame, while a frozen V-JEPA 2.1 temporal branch extracts clip-level egocentric context from the observed video. The temporal representation is injected into the detection pathway through feature modulation and ROI-level context fusion. The fused proposal features are then passed to multi-head STA predictors for box refinement, noun classification, verb classification, time-to-contact regression, and interaction confidence estimation. For the final submission, we further ensemble complementary predictions to improve robustness. Experimental results on the official challenge server show that VISTA achieves first place in the EgoVis 2026 Ego4D STA Challenge. Our code will be released at https://github.com/CorrineQiu/VISTA.

Qinglin Liu, Xiaoqian Lv, Quanling Meng et al.

Traditional studies emphasize the significance of context information in improving matting performance. Consequently, deep learning-based matting methods delve into designing pooling or affinity-based context aggregation modules to achieve superior results. However, these modules cannot well handle the context scale shift caused by the difference in image size during training and inference, resulting in matting performance degradation. In this paper, we revisit the context aggregation mechanisms of matting networks and find that a basic encoder-decoder network without any context aggregation modules can actually learn more universal context aggregation, thereby achieving higher matting performance compared to existing methods. Building on this insight, we present AEMatter, a matting network that is straightforward yet very effective. AEMatter adopts a Hybrid-Transformer backbone with appearance-enhanced axis-wise learning (AEAL) blocks to build a basic network with strong context aggregation learning capability. Furthermore, AEMatter leverages a large image training strategy to assist the network in learning context aggregation from data. Extensive experiments on five popular matting datasets demonstrate that the proposed AEMatter outperforms state-of-the-art matting methods by a large margin.

Harry Cheng, Yangyang Guo, Tianyi Wang et al.

The existing deepfake detection methods have reached a bottleneck in generalizing to unseen forgeries and manipulation approaches. Based on the observation that the deepfake detectors exhibit a preference for overfitting the specific primary regions in input, this paper enhances the generalization capability from a novel regularization perspective. This can be simply achieved by augmenting the images through primary region removal, thereby preventing the detector from over-relying on data bias. Our method consists of two stages, namely the static localization for primary region maps, as well as the dynamic exploitation of primary region masks. The proposed method can be seamlessly integrated into different backbones without affecting their inference efficiency. We conduct extensive experiments over three widely used deepfake datasets - DFDC, DF-1.0, and Celeb-DF with five backbones. Our method demonstrates an average performance improvement of 6% across different backbones and performs competitively with several state-of-the-art baselines.

Min Cao, Yang Bai, Jingyao Wang et al.

Under the flourishing development in performance, current image-text retrieval methods suffer from $N$-related time complexity, which hinders their application in practice. Targeting at efficiency improvement, this paper presents a simple and effective keyword-guided pre-screening framework for the image-text retrieval. Specifically, we convert the image and text data into the keywords and perform the keyword matching across modalities to exclude a large number of irrelevant gallery samples prior to the retrieval network. For the keyword prediction, we transfer it into a multi-label classification problem and propose a multi-task learning scheme by appending the multi-label classifiers to the image-text retrieval network to achieve a lightweight and high-performance keyword prediction. For the keyword matching, we introduce the inverted index in the search engine and create a win-win situation on both time and space complexities for the pre-screening. Extensive experiments on two widely-used datasets, i.e., Flickr30K and MS-COCO, verify the effectiveness of the proposed framework. The proposed framework equipped with only two embedding layers achieves $O(1)$ querying time complexity, while improving the retrieval efficiency and keeping its performance, when applied prior to the common image-text retrieval methods. Our code will be released.

Yisen Feng, Leigang Qu, Haoyu Zhang et al.

In this report, we present our champion solutions for the Natural Language Queries and GoalStep tracks of the Ego4D Episodic Memory Challenge at CVPR 2026. Both tracks require accurately localizing temporal segments from long untrimmed egocentric videos. To address these tasks, we propose a reranking-based framework that effectively leverages the strong video-language reasoning capability of multimodal large language model (MLLM) while preserving the efficiency and candidate recall of conventional localization pipelines. Specifically, we first obtain a set of candidate segments from existing localization model OSGNet, and then employ MLLM to select the segment that best matches the given query, thereby refining the final prediction. Ultimately, our method achieved first place in both the Natural Language Queries and GoalStep tracks. Our code can be found at https://github.com/iLearn-Lab/CVPR25-OSGNet.

Yangyang Guo, Haoyu Zhang, Yongkang Wong et al.

Learning a versatile language-image model is computationally prohibitive under a limited computing budget. This paper delves into the \emph{efficient language-image pre-training}, an area that has received relatively little attention despite its importance in reducing computational cost and footprint. To that end, we propose a vision token pruning and merging method ELIP, to remove less influential tokens based on the supervision of language outputs. Our method is designed with several strengths, such as being computation-efficient, memory-efficient, and trainable-parameter-free, and is distinguished from previous vision-only token pruning approaches by its alignment with task objectives. We implement this method in a progressively pruning manner using several sequential blocks. To evaluate its generalization performance, we apply ELIP to three commonly used language-image pre-training models and utilize public image-caption pairs with 4M images for pre-training. Our experiments demonstrate that with the removal of ~30$\%$ vision tokens across 12 ViT layers, ELIP maintains significantly comparable performance with baselines ($\sim$0.32 accuracy drop on average) over various downstream tasks including cross-modal retrieval, VQA, image captioning, \emph{etc}. In addition, the spared GPU resources by our ELIP allow us to scale up with larger batch sizes, thereby accelerating model pre-training and even sometimes enhancing downstream model performance.

Letian Shi, Qi Lv, Xiang Deng et al.

In this technical report, we present our solution for the EgoPlan Challenge in ICML 2024. To address the real-world egocentric task planning problem, we introduce a novel planning framework which comprises three stages: long-term memory Extraction, context-awared Planning, and multi-iteration Decision, named EPD. Given the task goal, task progress, and current observation, the extraction model first extracts task-relevant memory information from the progress video, transforming the complex long video into summarized memory information. The planning model then combines the context of the memory information with fine-grained visual information from the current observation to predict the next action. Finally, through multi-iteration decision-making, the decision model comprehensively understands the task situation and current state to make the most realistic planning decision. On the EgoPlan-Test set, EPD achieves a planning accuracy of 53.85% over 1,584 egocentric task planning questions. We have made all codes available at https://github.com/Kkskkkskr/EPD .

Yu-Wei Zhan, Fan Liu, Xin Luo et al.

Human-Object Interaction (HOI) detection aims at detecting human-object pairs and predicting their interactions. However, conventional HOI detection methods often struggle to fully capture the contextual information needed to accurately identify these interactions. While large Vision-Language Models (VLMs) show promise in tasks involving human interactions, they are not tailored for HOI detection. The complexity of human behavior and the diverse contexts in which these interactions occur make it further challenging. Contextual cues, such as the participants involved, body language, and the surrounding environment, play crucial roles in predicting these interactions, especially those that are unseen or ambiguous. Moreover, large VLMs are trained on vast image and text data, enabling them to generate contextual cues that help in understanding real-world contexts, object relationships, and typical interactions. Building on this, in this paper we introduce ConCue, a novel approach for improving visual feature extraction in HOI detection. Specifically, we first design specialized prompts to utilize large VLMs to generate contextual cues within an image. To fully leverage these cues, we develop a transformer-based feature extraction module with a multi-tower architecture that integrates contextual cues into both instance and interaction detectors. Extensive experiments and analyses demonstrate the effectiveness of using these contextual cues for HOI detection. The experimental results show that integrating ConCue with existing state-of-the-art methods significantly enhances their performance on two widely used datasets.

Junjie Wang, Xinghua Lou, Jason Li et al.

Text-to-Image (T2I) models and Unified Multimodal Models (UMMs) have achieved remarkable progress in visual generation. However, their reliance on a single-pass generation paradigm limits their ability to handle complex prompts requiring iterative refinement. To enable multi-round Reflective Visual Generation (RVG), we formalize the Reason-Reflect-Rectify (R^3) loop as a core framework and introduce R^3-Bench, a benchmark of over 600 expert-annotated instances that quantifies iterative reasoning and rectification capabilities. Evaluation on R^3-Bench reveals a critical gap: while state-of-the-art models can identify generation errors, they fail to generate actionable rectification instructions. To bridge this gap, we propose R^3-Refiner, a dual-stage framework leveraging Group Relative Policy Optimization (GRPO) and a Hierarchical Reward Mechanism (HRM) to better align rectification with reflective reasoning. Experiments show that R^3-Refiner achieves significant improvements on R^3-Bench (+12.0% in Reflective Verdict Score, +9.0% in Rectification Score), and can be seamlessly integrated with various MLLMs to enhance the generation quality of different T2I models on GenEval++ and T2I-CompBench. Code is available at https://github.com/xiaomoguhz/R3-Bench.

Haoyu Zhang, Qiaohui Chu, Yisen Feng et al.

This report presents MARS, short for Multimodal Agentic Reasoning with Source selection, our system for the CASTLE Challenge at EgoVis 2026. Participants must answer 185 closed-form questions over the CASTLE 2024 dataset. In contrast to prior single-video egocentric benchmarks, CASTLE requires reasoning over four days of activity, 15 synchronized perspectives, official transcripts, and multiple auxiliary modalities, including personal photos, auxiliary videos, gaze, thermal imagery, and heartrate measurements. MARS therefore treats the task as an agentic evidence-selection problem over multimodal sources rather than a purely text-only pipeline. MARS first follows the official CASTLE directory organization to build evidence memories from two primary sources, videos and transcripts, and four auxiliary sources, gaze, heartrate, photos, and thermal imagery. Long videos are converted into captions and DeepSeek-based summaries only because CASTLE videos are too long to fit directly into the model context for every question; this step compresses temporal evidence while keeping photos and other auxiliary media available as source-specific evidence. At inference time, a GPT-5.4 decision agent repeatedly chooses whether to continue reasoning, request a specific missing modality, produce an answer, or fall back to a random option when the evidence remains insufficient. The resulting system achieved second place on the final CASTLE Challenge leaderboard. Our codes are available at https://github.com/Hyu-Zhang/MARS.

Wei Li, Renshan Zhang, Rui Shao et al.

Vision-Language-Action (VLA) models have advanced in robotic manipulation, yet practical deployment remains hindered by two key limitations: 1) perceptual redundancy, where irrelevant visual inputs are processed inefficiently, and 2) superficial instruction-vision alignment, which hampers semantic grounding of actions. In this paper, we propose SemanticVLA, a novel VLA framework that performs Semantic-Aligned Sparsification and Enhancement for Efficient Robotic Manipulation. Specifically: 1) To sparsify redundant perception while preserving semantic alignment, Semantic-guided Dual Visual Pruner (SD-Pruner) performs: Instruction-driven Pruner (ID-Pruner) extracts global action cues and local semantic anchors in SigLIP; Spatial-aggregation Pruner (SA-Pruner) compacts geometry-rich features into task-adaptive tokens in DINOv2. 2) To exploit sparsified features and integrate semantics with spatial geometry, Semantic-complementary Hierarchical Fuser (SH-Fuser) fuses dense patches and sparse tokens across SigLIP and DINOv2 for coherent representation. 3) To enhance the transformation from perception to action, Semantic-conditioned Action Coupler (SA-Coupler) replaces the conventional observation-to-DoF approach, yielding more efficient and interpretable behavior modeling for manipulation tasks. Extensive experiments on simulation and real-world tasks show that SemanticVLA sets a new SOTA in both performance and efficiency. SemanticVLA surpasses OpenVLA on LIBERO benchmark by 21.1% in success rate, while reducing training cost and inference latency by 3.0-fold and 2.7-fold.SemanticVLA is open-sourced and publicly available at https://github.com/JiuTian-VL/SemanticVLA

Gongwei Chen, Leyang Shen, Rui Shao et al.

Multimodal Large Language Models (MLLMs) have endowed LLMs with the ability to perceive and understand multi-modal signals. However, most of the existing MLLMs mainly adopt vision encoders pretrained on coarsely aligned image-text pairs, leading to insufficient extraction and reasoning of visual knowledge. To address this issue, we devise a dual-Level vIsual knOwledge eNhanced Multimodal Large Language Model (LION), which empowers the MLLM by injecting visual knowledge in two levels. 1) Progressive incorporation of fine-grained spatial-aware visual knowledge. We design a vision aggregator cooperated with region-level vision-language (VL) tasks to incorporate fine-grained spatial-aware visual knowledge into the MLLM. To alleviate the conflict between image-level and region-level VL tasks during incorporation, we devise a dedicated stage-wise instruction-tuning strategy with mixture-of-adapters. This progressive incorporation scheme contributes to the mutual promotion between these two kinds of VL tasks. 2) Soft prompting of high-level semantic visual evidence. We facilitate the MLLM with high-level semantic visual evidence by leveraging diverse image tags. To mitigate the potential influence caused by imperfect predicted tags, we propose a soft prompting method by embedding a learnable token into the tailored text instruction. Comprehensive experiments on several multi-modal benchmarks demonstrate the superiority of our model (e.g., improvement of 5% accuracy on VSR and 3% CIDEr on TextCaps over InstructBLIP, 5% accuracy on RefCOCOg over Kosmos-2).

Zaijing Li, Yuquan Xie, Rui Shao et al.

Building a general-purpose agent is a long-standing vision in the field of artificial intelligence. Existing agents have made remarkable progress in many domains, yet they still struggle to complete long-horizon tasks in an open world. We attribute this to the lack of necessary world knowledge and multimodal experience that can guide agents through a variety of long-horizon tasks. In this paper, we propose a Hybrid Multimodal Memory module to address the above challenges. It 1) transforms knowledge into Hierarchical Directed Knowledge Graph that allows agents to explicitly represent and learn world knowledge, and 2) summarises historical information into Abstracted Multimodal Experience Pool that provide agents with rich references for in-context learning. On top of the Hybrid Multimodal Memory module, a multimodal agent, Optimus-1, is constructed with dedicated Knowledge-guided Planner and Experience-Driven Reflector, contributing to a better planning and reflection in the face of long-horizon tasks in Minecraft. Extensive experimental results show that Optimus-1 significantly outperforms all existing agents on challenging long-horizon task benchmarks, and exhibits near human-level performance on many tasks. In addition, we introduce various Multimodal Large Language Models (MLLMs) as the backbone of Optimus-1. Experimental results show that Optimus-1 exhibits strong generalization with the help of the Hybrid Multimodal Memory module, outperforming the GPT-4V baseline on many tasks.