CyberSec Research

Exams are a fundamental test of expert-level intelligence and require integrated understanding, reasoning, and generation. Existing exam-style benchmarks mainly focus on understanding and reasoning tasks, and current generation benchmarks emphasize the illustration of world knowledge and visual concepts, neglecting the evaluation of rigorous drawing exams. We introduce GenExam, the first benchmark for multidisciplinary text-to-image exams, featuring 1,000 samples across 10 subjects with exam-style prompts organized under a four-level taxonomy. Each problem is equipped with ground-truth images and fine-grained scoring points to enable a precise evaluation of semantic correctness and visual plausibility. Experiments show that even state-of-the-art models such as GPT-Image-1 and Gemini-2.5-Flash-Image achieve less than 15% strict scores, and most models yield almost 0%, suggesting the great challenge of our benchmark. By framing image generation as an exam, GenExam offers a rigorous assessment of models' ability to integrate knowledge, reasoning, and generation, providing insights on the path to general AGI.

While recent advancements in vision-language models have improved video understanding, diagnosing their capacity for deep, narrative comprehension remains a challenge. Existing benchmarks often test short-clip recognition or use template-based questions, leaving a critical gap in evaluating fine-grained reasoning over long-form narrative content. To address these gaps, we introduce $\mathsf{Cin\acute{e}aste}$, a comprehensive benchmark for long-form movie understanding. Our dataset comprises 3,119 multiple-choice question-answer pairs derived from 1,805 scenes across 200 diverse movies, spanning five novel fine-grained contextual reasoning categories. We use GPT-4o to generate diverse, context-rich questions by integrating visual descriptions, captions, scene titles, and summaries, which require deep narrative understanding. To ensure high-quality evaluation, our pipeline incorporates a two-stage filtering process: Context-Independence filtering ensures questions require video context, while Contextual Veracity filtering validates factual consistency against the movie content, mitigating hallucinations. Experiments show that existing MLLMs struggle on $\mathsf{Cin\acute{e}aste}$; our analysis reveals that long-range temporal reasoning is a primary bottleneck, with the top open-source model achieving only 63.15\% accuracy. This underscores significant challenges in fine-grained contextual understanding and the need for advancements in long-form movie comprehension.

High temporal resolution is essential for capturing fine-grained details in video understanding. However, current video large language models (VLLMs) and benchmarks mostly rely on low-frame-rate sampling, such as uniform sampling or keyframe selection, discarding dense temporal information. This compromise avoids the high cost of tokenizing every frame, which otherwise leads to redundant computation and linear token growth as video length increases. While this trade-off works for slowly changing content, it fails for tasks like lecture comprehension, where information appears in nearly every frame and requires precise temporal alignment. To address this gap, we introduce Dense Video Understanding (DVU), which enables high-FPS video comprehension by reducing both tokenization time and token overhead. Existing benchmarks are also limited, as their QA pairs focus on coarse content changes. We therefore propose DIVE (Dense Information Video Evaluation), the first benchmark designed for dense temporal reasoning. To make DVU practical, we present Gated Residual Tokenization (GRT), a two-stage framework: (1) Motion-Compensated Inter-Gated Tokenization uses pixel-level motion estimation to skip static regions during tokenization, achieving sub-linear growth in token count and compute. (2) Semantic-Scene Intra-Tokenization Merging fuses tokens across static regions within a scene, further reducing redundancy while preserving dynamic semantics. Experiments on DIVE show that GRT outperforms larger VLLM baselines and scales positively with FPS. These results highlight the importance of dense temporal information and demonstrate that GRT enables efficient, scalable high-FPS video understanding.

Recent progress in dense SLAM has primarily targeted monocular setups, often at the expense of robustness and geometric coverage. We present MCGS-SLAM, the first purely RGB-based multi-camera SLAM system built on 3D Gaussian Splatting (3DGS). Unlike prior methods relying on sparse maps or inertial data, MCGS-SLAM fuses dense RGB inputs from multiple viewpoints into a unified, continuously optimized Gaussian map. A multi-camera bundle adjustment (MCBA) jointly refines poses and depths via dense photometric and geometric residuals, while a scale consistency module enforces metric alignment across views using low-rank priors. The system supports RGB input and maintains real-time performance at large scale. Experiments on synthetic and real-world datasets show that MCGS-SLAM consistently yields accurate trajectories and photorealistic reconstructions, usually outperforming monocular baselines. Notably, the wide field of view from multi-camera input enables reconstruction of side-view regions that monocular setups miss, critical for safe autonomous operation. These results highlight the promise of multi-camera Gaussian Splatting SLAM for high-fidelity mapping in robotics and autonomous driving.

Vision Transformers (ViTs) achieve state-of-the-art performance in semantic segmentation but are hindered by high computational and memory costs. To address this, we propose STEP (SuperToken and Early-Pruning), a hybrid token-reduction framework that combines dynamic patch merging and token pruning to enhance efficiency without significantly compromising accuracy. At the core of STEP is dCTS, a lightweight CNN-based policy network that enables flexible merging into superpatches. Encoder blocks integrate also early-exits to remove high-confident supertokens, lowering computational load. We evaluate our method on high-resolution semantic segmentation benchmarks, including images up to 1024 x 1024, and show that when dCTS is applied alone, the token count can be reduced by a factor of 2.5 compared to the standard 16 x 16 pixel patching scheme. This yields a 2.6x reduction in computational cost and a 3.4x increase in throughput when using ViT-Large as the backbone. Applying the full STEP framework further improves efficiency, reaching up to a 4x reduction in computational complexity and a 1.7x gain in inference speed, with a maximum accuracy drop of no more than 2.0%. With the proposed STEP configurations, up to 40% of tokens can be confidently predicted and halted before reaching the final encoder layer.

Vision-centric Bird's Eye View (BEV) perception holds considerable promise for autonomous driving. Recent studies have prioritized efficiency or accuracy enhancements, yet the issue of domain shift has been overlooked, leading to substantial performance degradation upon transfer. We identify major domain gaps in real-world cross-domain scenarios and initiate the first effort to address the Domain Adaptation (DA) challenge in multi-view 3D object detection for BEV perception. Given the complexity of BEV perception approaches with their multiple components, domain shift accumulation across multi-geometric spaces (e.g., 2D, 3D Voxel, BEV) poses a significant challenge for BEV domain adaptation. In this paper, we introduce an innovative geometric-aware teacher-student framework, BEVUDA++, to diminish this issue, comprising a Reliable Depth Teacher (RDT) and a Geometric Consistent Student (GCS) model. Specifically, RDT effectively blends target LiDAR with dependable depth predictions to generate depth-aware information based on uncertainty estimation, enhancing the extraction of Voxel and BEV features that are essential for understanding the target domain. To collaboratively reduce the domain shift, GCS maps features from multiple spaces into a unified geometric embedding space, thereby narrowing the gap in data distribution between the two domains. Additionally, we introduce a novel Uncertainty-guided Exponential Moving Average (UEMA) to further reduce error accumulation due to domain shifts informed by previously obtained uncertainty guidance. To demonstrate the superiority of our proposed method, we execute comprehensive experiments in four cross-domain scenarios, securing state-of-the-art performance in BEV 3D object detection tasks, e.g., 12.9\% NDS and 9.5\% mAP enhancement on Day-Night adaptation.

Imagine living in a world composed solely of primitive shapes, could you still recognise familiar objects? Recent studies have shown that abstract images-constructed by primitive shapes-can indeed convey visual semantic information to deep learning models. However, representations obtained from such images often fall short compared to those derived from traditional raster images. In this paper, we study the reasons behind this performance gap and investigate how much high-level semantic content can be captured at different abstraction levels. To this end, we introduce the Hierarchical Abstraction Image Dataset (HAID), a novel data collection that comprises abstract images generated from normal raster images at multiple levels of abstraction. We then train and evaluate conventional vision systems on HAID across various tasks including classification, segmentation, and object detection, providing a comprehensive study between rasterised and abstract image representations. We also discuss if the abstract image can be considered as a potentially effective format for conveying visual semantic information and contributing to vision tasks.

This paper reviews the MARS2 2025 Challenge on Multimodal Reasoning. We aim to bring together different approaches in multimodal machine learning and LLMs via a large benchmark. We hope it better allows researchers to follow the state-of-the-art in this very dynamic area. Meanwhile, a growing number of testbeds have boosted the evolution of general-purpose large language models. Thus, this year's MARS2 focuses on real-world and specialized scenarios to broaden the multimodal reasoning applications of MLLMs. Our organizing team released two tailored datasets Lens and AdsQA as test sets, which support general reasoning in 12 daily scenarios and domain-specific reasoning in advertisement videos, respectively. We evaluated 40+ baselines that include both generalist MLLMs and task-specific models, and opened up three competition tracks, i.e., Visual Grounding in Real-world Scenarios (VG-RS), Visual Question Answering with Spatial Awareness (VQA-SA), and Visual Reasoning in Creative Advertisement Videos (VR-Ads). Finally, 76 teams from the renowned academic and industrial institutions have registered and 40+ valid submissions (out of 1200+) have been included in our ranking lists. Our datasets, code sets (40+ baselines and 15+ participants' methods), and rankings are publicly available on the MARS2 workshop website and our GitHub organization page https://github.com/mars2workshop/, where our updates and announcements of upcoming events will be continuously provided.

Digital beautification through social media filters has become increasingly popular, raising concerns about the reliability of facial images and videos and the effectiveness of automated face analysis. This issue is particularly critical for digital manipulation detectors, systems aiming at distinguishing between genuine and manipulated data, especially in cases involving deepfakes and morphing attacks designed to deceive humans and automated facial recognition. This study examines whether beauty filters impact the performance of deepfake and morphing attack detectors. We perform a comprehensive analysis, evaluating multiple state-of-the-art detectors on benchmark datasets before and after applying various smoothing filters. Our findings reveal performance degradation, highlighting vulnerabilities introduced by facial enhancements and underscoring the need for robust detection models resilient to such alterations.

Accurate histopathological diagnosis often requires multiple differently stained tissue sections, a process that is time-consuming, labor-intensive, and environmentally taxing due to the use of multiple chemical stains. Recently, virtual staining has emerged as a promising alternative that is faster, tissue-conserving, and environmentally friendly. However, existing virtual staining methods face significant challenges in clinical applications, primarily due to their reliance on well-aligned paired data. Obtaining such data is inherently difficult because chemical staining processes can distort tissue structures, and a single tissue section cannot undergo multiple staining procedures without damage or loss of information. As a result, most available virtual staining datasets are either unpaired or roughly paired, making it difficult for existing methods to achieve accurate pixel-level supervision. To address this challenge, we propose a robust virtual staining framework featuring cascaded registration mechanisms to resolve spatial mismatches between generated outputs and their corresponding ground truth. Experimental results demonstrate that our method significantly outperforms state-of-the-art models across five datasets, achieving an average improvement of 3.2% on internal datasets and 10.1% on external datasets. Moreover, in datasets with substantial misalignment, our approach achieves a remarkable 23.8% improvement in peak signal-to-noise ratio compared to baseline models. The exceptional robustness of the proposed method across diverse datasets simplifies the data acquisition process for virtual staining and offers new insights for advancing its development.

Self-supervised learning through masked autoencoders has attracted great attention for remote sensing (RS) foundation model (FM) development, enabling improved representation learning across diverse sensors and downstream tasks. However, existing RS FMs often either suffer from substantial computational complexity during both training and inference or exhibit limited representational capacity. These issues restrict their practical applicability in RS. To address this limitation, we propose an adaptation for enhancing the efficiency of RS FMs by integrating the Soft mixture-of-experts (MoE) mechanism into the FM. The integration of Soft MoEs into the FM allows modality-specific expert specialization alongside shared cross-sensor representation learning. To demonstrate the effectiveness of our adaptation, we apply it on the Cross-Sensor Masked Autoencoder (CSMAE) model, resulting in the Cross-Sensor Mixture-of-Experts (CSMoE) model. In addition, we introduce a thematic-climatic descriptor-driven sampling strategy for the construction of a representative and diverse training set to train our CSMoE model. Extensive experiments on scene classification, semantic segmentation, and content-based image retrieval demonstrate that our adaptation yields a reduction in computational requirements while maintaining or improving representational performance. Compared to state-of-the-art RS FMs, CSMoE achieves a superior trade-off between representational capacity, accuracy, and computational efficiency. On average, CSMoE achieves more than twice the computational efficiency of existing RS FMs, while maintaining competitive performance across all experiments. These results show the effectiveness of the proposed adaptation for creating computationally efficient RS FMs. The code for the model, the training set creation, and the model weights will be available at https://git.tu-berlin.de/rsim/csmoe.

Weakly-supervised audio-visual video parsing (AVVP) seeks to detect audible, visible, and audio-visual events without temporal annotations. Previous work has emphasized refining global predictions through contrastive or collaborative learning, but neglected stable segment-level supervision and class-aware cross-modal alignment. To address this, we propose two strategies: (1) an exponential moving average (EMA)-guided pseudo supervision framework that generates reliable segment-level masks via adaptive thresholds or top-k selection, offering stable temporal guidance beyond video-level labels; and (2) a class-aware cross-modal agreement (CMA) loss that aligns audio and visual embeddings at reliable segment-class pairs, ensuring consistency across modalities while preserving temporal structure. Evaluations on LLP and UnAV-100 datasets shows that our method achieves state-of-the-art (SOTA) performance across multiple metrics.

Zero-Shot Anomaly Detection (ZSAD) seeks to identify anomalies from arbitrary novel categories, offering a scalable and annotation-efficient solution. Traditionally, most ZSAD works have been based on the CLIP model, which performs anomaly detection by calculating the similarity between visual and text embeddings. Recently, vision foundation models such as DINOv3 have demonstrated strong transferable representation capabilities. In this work, we are the first to adapt DINOv3 for ZSAD. However, this adaptation presents two key challenges: (i) the domain bias between large-scale pretraining data and anomaly detection tasks leads to feature misalignment; and (ii) the inherent bias toward global semantics in pretrained representations often leads to subtle anomalies being misinterpreted as part of the normal foreground objects, rather than being distinguished as abnormal regions. To overcome these challenges, we introduce AD-DINOv3, a novel vision-language multimodal framework designed for ZSAD. Specifically, we formulate anomaly detection as a multimodal contrastive learning problem, where DINOv3 is employed as the visual backbone to extract patch tokens and a CLS token, and the CLIP text encoder provides embeddings for both normal and abnormal prompts. To bridge the domain gap, lightweight adapters are introduced in both modalities, enabling their representations to be recalibrated for the anomaly detection task. Beyond this baseline alignment, we further design an Anomaly-Aware Calibration Module (AACM), which explicitly guides the CLS token to attend to anomalous regions rather than generic foreground semantics, thereby enhancing discriminability. Extensive experiments on eight industrial and medical benchmarks demonstrate that AD-DINOv3 consistently matches or surpasses state-of-the-art methods, verifying its superiority as a general zero-shot anomaly detection framework.

Current multi-object tracking (MOT) algorithms typically overlook issues inherent in low-quality videos, leading to significant degradation in tracking performance when confronted with real-world image deterioration. Therefore, advancing the application of MOT algorithms in real-world low-quality video scenarios represents a critical and meaningful endeavor. To address the challenges posed by low-quality scenarios, inspired by vision-language models, this paper proposes a Visual Semantic Enhancement-guided Multi-Object Tracking framework (VSE-MOT). Specifically, we first design a tri-branch architecture that leverages a vision-language model to extract global visual semantic information from images and fuse it with query vectors. Subsequently, to further enhance the utilization of visual semantic information, we introduce the Multi-Object Tracking Adapter (MOT-Adapter) and the Visual Semantic Fusion Module (VSFM). The MOT-Adapter adapts the extracted global visual semantic information to suit multi-object tracking tasks, while the VSFM improves the efficacy of feature fusion. Through extensive experiments, we validate the effectiveness and superiority of the proposed method in real-world low-quality video scenarios. Its tracking performance metrics outperform those of existing methods by approximately 8% to 20%, while maintaining robust performance in conventional scenarios.

We introduce Wan-Animate, a unified framework for character animation and replacement. Given a character image and a reference video, Wan-Animate can animate the character by precisely replicating the expressions and movements of the character in the video to generate high-fidelity character videos. Alternatively, it can integrate the animated character into the reference video to replace the original character, replicating the scene's lighting and color tone to achieve seamless environmental integration. Wan-Animate is built upon the Wan model. To adapt it for character animation tasks, we employ a modified input paradigm to differentiate between reference conditions and regions for generation. This design unifies multiple tasks into a common symbolic representation. We use spatially-aligned skeleton signals to replicate body motion and implicit facial features extracted from source images to reenact expressions, enabling the generation of character videos with high controllability and expressiveness. Furthermore, to enhance environmental integration during character replacement, we develop an auxiliary Relighting LoRA. This module preserves the character's appearance consistency while applying the appropriate environmental lighting and color tone. Experimental results demonstrate that Wan-Animate achieves state-of-the-art performance. We are committed to open-sourcing the model weights and its source code.

Almost 30% of prostate cancer (PCa) patients undergoing radical prostatectomy (RP) experience biochemical recurrence (BCR), characterized by increased prostate specific antigen (PSA) and associated with increased mortality. Accurate early prediction of BCR, at the time of RP, would contribute to prompt adaptive clinical decision-making and improved patient outcomes. In this work, we propose prostate cancer BCR prediction via fused multi-modal embeddings (PROFUSEme), which learns cross-modal interactions of clinical, radiology, and pathology data, following an intermediate fusion configuration in combination with Cox Proportional Hazard regressors. Quantitative evaluation of our proposed approach reveals superior performance, when compared with late fusion configurations, yielding a mean C-index of 0.861 ($\sigma=0.112$) on the internal 5-fold nested cross-validation framework, and a C-index of 0.7103 on the hold out data of CHIMERA 2025 challenge validation leaderboard.

We introduce SAIL-VL2, an open-suite vision-language foundation model (LVM) for comprehensive multimodal understanding and reasoning. As the successor to SAIL-VL, SAIL-VL2 achieves state-of-the-art performance at the 2B and 8B parameter scales across diverse image and video benchmarks, demonstrating strong capabilities from fine-grained perception to complex reasoning. Three core innovations drive its effectiveness. First, a large-scale data curation pipeline with scoring and filtering strategies enhances both quality and distribution across captioning, OCR, QA, and video data, improving training efficiency. Second, a progressive training framework begins with a powerful pre-trained vision encoder (SAIL-ViT), advances through multimodal pre-training, and culminates in a thinking-fusion SFT-RL hybrid paradigm that systematically strengthens model capabilities. Third, architectural advances extend beyond dense LLMs to efficient sparse Mixture-of-Experts (MoE) designs. With these contributions, SAIL-VL2 demonstrates competitive performance across 106 datasets and achieves state-of-the-art results on challenging reasoning benchmarks such as MMMU and MathVista. Furthermore, on the OpenCompass leaderboard, SAIL-VL2-2B ranks first among officially released open-source models under the 4B parameter scale, while serving as an efficient and extensible foundation for the open-source multimodal community.

Drone detection in visually complex environments remains challenging due to background clutter, small object scale, and camouflage effects. While generic object detectors like YOLO exhibit strong performance in low-texture scenes, their effectiveness degrades in cluttered environments with low object-background separability. To address these limitations, this work presents an enhanced iteration of YOLO-FEDER FusionNet -- a detection framework that integrates generic object detection with camouflage object detection techniques. Building upon the original architecture, the proposed iteration introduces systematic advancements in training data composition, feature fusion strategies, and backbone design. Specifically, the training process leverages large-scale, photo-realistic synthetic data, complemented by a small set of real-world samples, to enhance robustness under visually complex conditions. The contribution of intermediate multi-scale FEDER features is systematically evaluated, and detection performance is comprehensively benchmarked across multiple YOLO-based backbone configurations. Empirical results indicate that integrating intermediate FEDER features, in combination with backbone upgrades, contributes to notable performance improvements. In the most promising configuration -- YOLO-FEDER FusionNet with a YOLOv8l backbone and FEDER features derived from the DWD module -- these enhancements lead to a FNR reduction of up to 39.1 percentage points and a mAP increase of up to 62.8 percentage points at an IoU threshold of 0.5, compared to the initial baseline.