Jiajun Zhang, Yuying Li, Zhixun Li, Xingyu Guo, Jingzhuo Wu, Leqi Zheng, Yiran Yang, Jianke Zhang, Qingbin Li, Shannan Yan, Changguo Jia, Junfei Wu, Zilei Wang, Qiang Liu, Liang Wang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Junfei Wu, Yue Ding, Guofan Liu, Tianze Xia, Ziyue Huang, Dianbo Sui, Qiang Liu, Shu Wu, Liang Wang, Tieniu Tan. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Comprehensive evaluation of Multimodal Large Language Models (MLLMs) has recently garnered widespread attention in the research community. However, we observe that existing benchmarks present several common barriers that make it difficult to measure the significant challenges that models face in the real world, including: 1) small data scale leads to a large performance variance; 2) reliance on model-based annotations results in restricted data quality; 3) insufficient task difficulty, especially caused by the limited image resolution. To tackle these issues, we introduce MME-RealWorld. Specifically, we collect more than $300$ K images from public datasets and the Internet, filtering $13,366$ high-quality images for annotation. This involves the efforts of professional $25$ annotators and $7$ experts in MLLMs, contributing to $29,429$ question-answer pairs that cover $43$ subtasks across $5$ real-world scenarios, extremely challenging even for humans. As far as we know, **MME-RealWorld is the largest manually annotated benchmark to date, featuring the highest resolution and a targeted focus on real-world applications**. We further conduct a thorough evaluation involving $29$ prominent MLLMs, such as GPT-4o, Gemini 1.5 Pro, and Claude 3.5 Sonnet. Our results show that even the most advanced models struggle with our benchmarks, where none of them reach 60\% accuracy. The challenges of perceiving high-resolution images and understanding complex real-world scenarios remain urgent issues to be addressed. The data and evaluation code are released in our Project Page.

Inspired by DeepSeek-R1's success in eliciting reasoning abilities through rule-based reinforcement learning (RL), we introduce Video-R1 as the first attempt to systematically explore the R1 paradigm for incentivizing video reasoning within multimodal large language models (MLLMs). However, directly applying RL training with the GRPO algorithm to video reasoning presents two primary challenges: (i) a lack of temporal modeling for video reasoning, and (ii) the scarcity of high-quality video-reasoning data. To address these issues, we first propose the T-GRPO algorithm, which encourages models to utilize temporal information in videos for reasoning. Additionally, instead of relying solely on video data, we incorporate high-quality image-reasoning data into the training process. We have constructed two datasets: Video-R1-CoT-165k for SFT cold start and Video-R1-260k for RL training, both comprising image and video data. Experimental results demonstrate that Video-R1 achieves significant improvements on video reasoning benchmarks such as VideoMMMU and VSI-Bench, as well as on general video benchmarks including MVBench and TempCompass, etc. Notably, Video-R1-7B attains a 37.1\% accuracy on video spatial reasoning benchmark VSI-bench, surpassing the commercial proprietary model GPT-4o. All code, models, and data will be released.

As textual reasoning with large language models (LLMs) has advanced significant, there has been growing interest in enhancing the multimodal reasoning capabilities of large vision-language models (LVLMs). However, existing methods primarily approach multimodal reasoning in a straightforward, text-centric manner, where both reasoning and answer derivation are conducted purely through text, with the only difference being the presence of multimodal input. As a result, these methods often encounter fundamental limitations in spatial reasoning tasks that demand precise geometric understanding and continuous spatial tracking\textemdash capabilities that humans achieve through mental visualization and manipulation. To address the limitations, we propose drawing to reason in space, a novel paradigm that enables LVLMs to reason through elementary drawing operations in the visual space. By equipping models with basic drawing operations including annotating bounding boxes and drawing auxiliary lines, we empower them to express and analyze spatial relationships through direct visual manipulation, meanwhile avoiding the performance ceiling imposed by specialized perception tools in previous tool-integrated reasoning approaches. To cultivate this capability, we develop a three-stage training framework: cold-start training with synthetic data to establish basic drawing abilities, reflective rejection sampling to enhance self-reflection behaviors, and reinforcement learning to directly optimize for target rewards. Extensive experiments demonstrate that our model, named \textsc{Spark}, consistently outperforms existing methods across diverse spatial reasoning benchmarks involving maze navigation, static spatial reasoning, video-based reasoning and multi-view-based reasoning tasks, with an average improvement of 11.5\%. Ablation studies reveal the critical role of each training stage, with reflective rejection sampling particularly enhancing the model's self-correction capabilities and reasoning potential.

Evidence-aware fake news detection aims to conduct reasoning between news and evidences, which are retrieved based on news content, to find uniformity or inconsistency. However, we find evidence-aware detection models suffer from biases, i.e., spurious correlations between news/evidence contents and true/fake news labels, and are hard to be generalized to Out-Of-Distribution (OOD) situations. To deal with this, we propose a novel Dual Adversarial Learning (DAL) approach. We incorporate news-aspect and evidence-aspect debiasing discriminators, whose targets are both true/fake news labels, in DAL. Then, DAL reversely optimizes news-aspect and evidence-aspect debiasing discriminators to mitigate the impact of news and evidence content biases. At the same time, DAL also optimizes the main fake news predictor, so that the news-evidence interaction module can be learned. This process allows us to teach evidence-aware fake news detection models to better conduct news-evidence reasoning, and minimize the impact of content biases. To be noted, our proposed DAL approach is a plug-and-play module that works well with existing backbones. We conduct comprehensive experiments under two OOD settings, and plug DAL in four evidence-aware fake news detection backbones. Results demonstrate that, DAL significantly and stably outperforms the original backbones and some competitive debiasing methods.

With the rapid expansion of multimodal content in online social media, automatic detection of multimodal fake news has received much attention. Multimodal joint training commonly used in existing methods is expected to benefit from thoroughly leveraging cross-modal features, yet these methods still suffer from insufficient learning of uni-modal features. Due to the heterogeneity of multimodal networks, optimizing a single objective will inevitably make the models prone to rely on specific modality while leaving other modalities under-optimized. On the other hand, simply expecting each modality to play a significant role in identifying all the rumors is also not appropriate as the multimodal fake news often involves tampering in only one modality. Therefore, how to find the genuine tampering on the per-sample basis becomes the key point to unlock the full power of each modality in a good collaborative manner. To address these issues, we propose aUni-modalEvent-agnosticKnowledgeDistillation framework (UEKD), which aims to transfer knowledge contained in the fine-grained prediction from uni-modal teachers to the multimodal student model through modality-specific distillation. Specifically, we find that the uni-modal teachers simply trained on the whole training set are easy to memorize the event-specific noise information to make a correct but biased prediction, failing to reflect the genuine degree of tampering in each modality. To tackle this problem, we propose to train and validate the teacher models on different domains in training dataset through a cross-validation manner, as the predictions from the out-of-domain teachers can be regarded as event-agnostic knowledge without spurious connections with event-specific information. Finally, to balance the convergence speeds across modalities, we dynamically monitor the involvement of each modality during training, through which we could identify the more under-optimized modalities and re-weight the distillation loss accordingly. Our method could be served as a plug-and-play module for existing multimodal fake news detection backbones. Extensive experiments on three public datasets and four state-of-the-art fake news detection backbones show that our proposed method can improve the performance by a large margin.

The prevalence and perniciousness of fake news have been a critical issue on the Internet, which stimulates the development of automatic fake news detection in turn. In this paper, we focus on the evidence-based fake news detection, where several evidences are utilized to probe the veracity of news (i.e., a claim). Most previous methods first employ sequential models to embed the semantic information and then capture the claim-evidence interaction based on different attention mechanisms. Despite their effectiveness, they still suffer from three weaknesses. Firstly, due to the inherent drawbacks of sequential models, they fail to integrate the relevant information that is scattered far apart in evidences for veracity checking. Secondly, they underestimate much redundant information contained in evidences that may be useless or even harmful. Thirdly, insufficient data utilization limits the separability and reliability of representations captured by the model, which are sensitive to local evidence. To solve these problems, we propose a unifiedGraph-based sEmantic structure mining framework with ConTRAstiveLearning, namely GETRAL in short. Specifically, different from the existing work that treats claims and evidences as sequences, we first model them as graph-structured data and capture the long-distance semantic dependency among dispersed relevant snippets via neighborhood propagation. After obtaining contextual semantic information, our model reduces information redundancy by performing graph structure learning. Then the fine-grained semantic representations are fed into the downstream claim-evidence interaction module for predictions. Finally, the supervised contrastive learning accompanied with adversarial augmented instances is applied to make full use of data and strengthen the representation learning. Comprehensive experiments have demonstrated the superiority of GETRAL over the state-of-the-arts and validated the efficacy of semantic mining with graph structure and contrastive learning.

Evidence-based fake news detection is to judge the veracity of news against relevant evidences. However, models tend to memorize the dataset biases within spurious correlations between news patterns and veracity labels as shortcuts, rather than learning how to integrate the information behind them to reason. As a consequence, models may suffer from a serious failure when facing real-life conditions where most news has different patterns. Inspired by the success of causal inference, we propose a novel framework for debiasing evidence-based fake news detection\footnoteCode available at https://github.com/CRIPAC-DIG/CF-FEND by causal intervention. Under this framework, the model is first trained on the original biased dataset like ordinary work, then it makes conventional predictions and counterfactual predictions simultaneously in the testing stage, where counterfactual predictions are based on the intervened evidence. Relatively unbiased predictions are obtained by subtracting intervened outputs from the conventional ones. Extensive experiments conducted on several datasets demonstrate our method's effectiveness and generality on debiased datasets.

The prevalence and perniciousness of fake news has been a critical issue on the Internet, which stimulates the development of automatic fake news detection in turn. In this paper, we focus on the evidence-based fake news detection, where several evidences are utilized to probe the veracity of news (i.e., a claim). Most previous methods first employ sequential models to embed the semantic information and then capture the claim-evidence interaction based on different attention mechanisms. Despite their effectiveness, they still suffer from two main weaknesses. Firstly, due to the inherent drawbacks of sequential models, they fail to integrate the relevant information that is scattered far apart in evidences for veracity checking. Secondly, they neglect much redundant information contained in evidences that may be useless or even harmful. To solve these problems, we propose a unified Graph-based sEmantic sTructure mining framework, namely GET in short. Specifically, different from the existing work that treats claims and evidences as sequences, we model them as graph-structured data and capture the long-distance semantic dependency among dispersed relevant snippets via neighborhood propagation. After obtaining contextual semantic information, our model reduces information redundancy by performing graph structure learning. Finally, the fine-grained semantic representations are fed into the downstream claim-evidence interaction module for predictions. Comprehensive experiments have demonstrated the superiority of GET over the state-of-the-arts.

Visual tracking is still a challenging fundamental task in the field of computer vision, especially in complex scenes such as long-term occlusion, nonrigid deformation and fast movement. In this paper, we presented an RGB-D tracker based on the Siamese Region Proposal Network and Depth Information. First, Siamese Network with shared parameters was constructed to perform feature extraction on the target patch and search area. Second, Region Proposal Network was constructed to estimate the target position in the RGB channels. At the same time, the depth information in the RGB-D video was used to determine the target occlusion state and fine-tune the target position. Finally, the tracker used depth information to achieve occlusion recovery when the target was fully occluded. The experimental result shows that the method has better performance in tracking accuracy and tracking speed on the large-scale Princeton RGB-D Tracking Benchmark (PTB) dataset.