Multimodal large language models (MLLMs), which integrate language and visual cues for problem-solving, are crucial for advancing artificial general intelligence (AGI). However, current benchmarks for measuring the intelligence of MLLMs suffer from limited scale, narrow coverage, and unstructured knowledge, offering only static and undifferentiated evaluations. To bridge this gap, we introduce MDK12-Bench, a large-scale multidisciplinary benchmark built from real-world K–12 exams spanning six disciplines with 141K instances and 6,225 knowledge points organized in a six-layer taxonomy. Covering five question formats with difficulty and year annotations, it enables comprehensive evaluation to capture the extent to which MLLMs perform over four dimensions: 1) difficulty levels, 2) temporal (cross-year) shifts, 3) contextual shifts, and 4) knowledge-driven reasoning. We propose a novel dynamic evaluation framework that introduces unfamiliar visual, textual, and question form shifts to challenge model generalization while improving benchmark objectivity and longevity by mitigating data contamination. We further evaluate knowledge-point reference-augmented generation (KP-RAG) to examine the role of knowledge in reasoning. Key findings reveal limitations in current MLLMs in multiple aspects and provide guidance for enhancing model reasoning, robustness, and AI-assisted education.

Zizhen Li, Chuanhao Li, Yibin Wang, Jianwen Sun, Yukang Feng, Jiaxin Ai, Fanrui Zhang, Mingzhu Sun, Yifei Huang, Kaipeng Zhang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Recent studies show that claims incorporating both text and images spread more effectively than those with text alone, presenting significant challenges for multimodal fact-checking. The rapid development of Multi-modal Large Language Models (MLLMs) has greatly advanced research in this field, enabling stronger performance. However, existing MLLM-based fact-checking methods fail to fully exploit visual evidence, and their reliance on rigid fine-tuning templates limits context-aware explanations and leads to weak deep reasoning. To address these limitations, we propose FACTCOMPASS, a novel framework that combines reasoning-aware fine-tuning with large-scale rule-based reinforcement learning and incorporates a semantic- and knowledge-enhanced retrieval module to strengthen deep reasoning and improve evidence utilization. This framework enhances evidence retrieval by obtaining semantically relevant evidence images, enriching the contextual understanding of claim-related images, and refining textual evidence at the knowledge level. To further enhance reasoning, we introduce a self-refining reinforcement fine-tuning strategy: (1) distilling GPT-4o's reasoning from partially fact-checking data for cold-start Chain-of-Thought learning; (2) activating reasoning across broader datasets using prior knowledge and rejection sampling; (3) applying Group Relative Policy Optimization to explore diverse reasoning paths and optimize factual consistency. Extensive experiments have demonstrated the effectiveness of the proposed framework.

Zizhen Li, Chuanhao Li, Yibin Wang, Qi Chen, Diping Song, Yukang Feng, Jianwen Sun, Jiaxin Ai, Fanrui Zhang, Mingzhu Sun, Kaipeng Zhang. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Video generation techniques have made remarkable progress, promising to be the foundation of interactive world exploration.However, existing video generation datasets are not well-suited for world exploration training as they suffer from some limitations: limited locations, short duration, static scenes, and a lack of annotations about exploration and the world.In this paper, we introduce Sekai (meaning "world" in Japanese), a high-quality first-person view worldwide video dataset with rich annotations for world exploration. It consists of over 5,000 hours of walking or drone view (FPV and UVA) videos from over 100 countries and regions across 750 cities. We develop an efficient and effective toolbox to collect, pre-process and annotate videos with location, scene, weather, crowd density, captions, and camera trajectories.Comprehensive analyses and experiments demonstrate the dataset’s scale, diversity, annotation quality, and effectiveness for training video generation models.We believe Sekai will benefit the area of video generation and world exploration, and motivate valuable applications.

The absence of temporal information in single-temporal satellite remote sensing images presents a substantial challenge for target state discrimination. In this letter, a pioneering Remote Sensing Airplane State Discrimination Network (RSASDNet) is introduced, by leveraging the relationship between targets and their backgrounds in single-temporal high-resolution remote sensing images. To facilitate the study, we take airplane state discrimination as an example, and a Remote Sensing Airport Panoptic Segmentation with Airplane States Dataset (RSAPS-ASD) is constructed. RSASDNet incorporates two key innovations: 1) a scene knowledge graph generation module that constructs scene knowledge representation by capturing spatial relationships between airplane instances and their surrounding environment (e.g., taxiways and hangars); and 2) a novel graph-image hybrid convolution discrimination module that synergistically integrates structural knowledge and spatial semantic information through dedicated dual-branch learning. The effectiveness of the proposed method is validated using RSAPS-ASD, with experimental results demonstrating that RSASDNet achieves an impressive accuracy of 73.95% in airplane state discrimination.

In vivo cell manipulation is crucial for understanding organ functions and dysfunctions at the cellular level. Due to its low optical transparency, the in vivo environment needs to be lighted up first to visualize the targets for manipulation. At present, operators usually carefully blow a micropipette filled with fluorescent dye solution to create a fluorescence field around the micropipette opening, lighting up the in vivo environment. This manual dynamic control of the fluorescence field is usually a low-accuracy, labor-intensive, and high-skill requirement task, becoming even more challenging when the micropipette moves to locate or approach target cells. A large number of in vivo lighting-control tests may improve manual control efficiency of the fluorescent field, but they usually cost many precise animal samples, making them costly and often unaffordable. To enhance the dynamic control efficiency of the fluorescent field while reducing experimental costs, a robotic fluorescent lighting control method is proposed based on an in vivo dynamic fluorescent imaging simulator. First, a dynamic fluorescence imaging model composing an injection model, a diffusion model, and a luminescence model of fluorescent molecules is established to simulate the fluorescence field in vivo. The fluorescent intensity distribution obtained through the model is highly in accordance with experimental results. Based on this model, an adaptive sliding mode controller is employed to achieve the desired fluorescent intensity in the brain tissue. Both the simulation results and the experimental results demonstrate significant advantages of the proposed controller in terms of control accuracy and stability compared to the traditional PID controller and manual operation. Note to Practitioners—Forming a stable fluorescent light field with a specified fluorescent intensity is of vital importance operation for in vivo micromanipulations. At present, manual adjustment of the fluorescent field by blowing the micropipette with mouth is time-consuming and energy-draining. In this paper, we first propose a dynamic fluorescence imaging simulator to address the economic and time costs of live experiments. We then design an adaptive sliding mode controller to dynamically control the fluorescent intensity at a specified position in the fluorescent field. Simulations based on the established model, along with experiments on brain tissues, demonstrate the controller’s effectiveness. Our system is expected to reduce human involvement and enable high-precision lighting for in vivo cell manipulation in the future.

Accurate estimation of channel state information (CSI) is crucial to realize the full potential of massive multiple-input multiple-output (MIMO) communication systems. Despite the high accuracy of existing CSI estimators, the high computational complexity makes them impractical for real-world massive MIMO systems. So, developing a highly accurate and low-complexity estimator has been a long-standing challenge in the field of massive MIMO. In this paper, we present a channel estimation scheme that significantly reduces computational complexity while achieving high accuracy. Firstly, we design a two-stage training protocol whereby only part of the transmit and receive antennas are activated for signal emission or reception. Secondly, we apply the low-complexity Least-Square (LS) estimator to acquire two sub-blocks of the channel matrix. Relying on the inherent low-rank property of channel matrix, the complete channel matrix is reconstructed via a randomized matrix approximation technique. Thirdly, we consider three different power allocation schemes to further optimize the pilot power in 2-stage training process. The theoretical bounds of estimation error for our CSI estimator are derived, and the optimal power allocation strategy is thus obtained by minimizing this error bound. Numerical simulations are provided to demonstrate our proposed method. As shown, the theoretical error bound is tight, and the optimal power allocation can achieve the substantial gain. Our CSI estimator incurs the same complexity as the popular LS estimator, whilst the estimation accuracy is improved by ~5 dB, which has the great promise to new-generation massive MIMO communications.

Two-photon microscopy is currently the preferred technique for in vivo vasculature imaging. The ability to generate high-resolution 3D images of blood vessels could improve our understanding of normal vascular physiology and disease pathogenesis. Automatic segmenting and mapping of these 3D blood vessel networks remain challenging due to vascular geometries' complexity and signal-to-noise ratio limitations. In this paper, we propose a novel segmentation network named HYbrid Swin Unet Transformers (HiSwin UT) to address this problem. The network explores the use of the Swin transformer to segment 3D volumetric vessels in vivo images acquired by two-photon microscopy. We propose to use a hybrid of hierarchical Swin transformers with 3D convolutional neural networks (3D CNNs) as both the encoder and the decoder. The Swin transformer guards guide the 3D CNN laborers at every stage in both the encoding and decoding paths to enlarge the receptive field of the network. Specifically, the 3D vasculature images are projected into a sequence of embedding and fed into the network. The encoders extract features at five different resolutions by using a combination of sifted windows self-attention layers and residual convolutional layers. Then these latent multi-scale features are connected to each decoder via skip connection to cater to the diverse range of vessels in terms of their sizes and shapes. We evaluate this HiSwin UT model on the public two-photon vasculature dataset, extensive experimental results demonstrate the superiority of our network compared with the state-of-the-art models. The source code of the HiSwin UT will be publicly available at https://github.com/UTINK/HiSwinUT.