arXiv cs.AI
·
6 hours ago
Researchers developed a zero-setup framework using a pretrained semantic segmentation network to automatically segment X-ray tomography images without manual thresholding or retraining for each new dataset. The framework generates segmentation masks within minutes of reconstruction and works across previously unseen datasets without user input. This enables rapid assessment of scan quality and material properties during ongoing experiments, supporting real-time beamline feedback and scalable scientific imaging workflows.
arXiv cs.AI
·
6 hours ago
Researchers developed FlowWM, a stochastic world model that uses flow matching directly within pretrained feature spaces like DINOv3 to forecast uncertain futures while maintaining information for downstream perception tasks. The model introduces a differentiable one-step projection mechanism to handle high-dimensional features efficiently, and evaluation on synthetic and real-world FuturePerception benchmarks showed improvements in perception performance, mode coverage, and horizon robustness compared to existing approaches. The method addresses limitations of previous visual world models by avoiding both the low-dimensional collapse of VAE-based approaches and the multimodal future collapse of deterministic predictors.
arXiv cs.AI
·
6 hours ago
Researchers introduced Concept-Guided In-Context Segmentation (CG-ICS), a method that improves robustness in image segmentation tasks by using semantic concepts extracted from reference images rather than relying on visual matching alone. The approach combines an MLLM-based concept reasoning module with SAM3 backbone to produce stable segmentation results across different reference images. The method achieves state-of-the-art accuracy while substantially reducing variance in segmentation outputs when different reference sets are provided.
arXiv cs.AI
·
6 hours ago
Conan-embedding-v3 is a framework for omni-modal retrieval that trains separate specialist models for text, image, video, document, and audio before fusing their task vectors into a single backbone. The model achieves 74.9 scores on MMEB and 55.61 on the 30-task MAEB audio suite. The framework addresses a projector drift problem where audio retrieval degrades after fusion by applying projector recovery and balanced multi-modal rehearsal to maintain performance across all modalities.
arXiv cs.AI
·
6 hours ago
Researchers developed REST, a training-free framework for zero-shot object-goal navigation that represents possible movement options as a tree of paths rather than individual waypoints, allowing language models to reason about routes more efficiently. The system builds a 3D map from RGB-D camera streams, generates candidate paths through sampling-based planning, and uses language model reasoning to select the next-best route, achieving top performance on Gibson, HM3D, and HSSD benchmarks. This approach enables the system to discover target objects in unfamiliar environments without task-specific training by considering information gathered along routes rather than just at destinations.
arXiv cs.AI
·
6 hours ago
Researchers developed QARMVC, a deep learning framework for multi-view clustering that handles varying levels of noise across different data points by using reconstruction discrepancy to measure contamination intensity and quality scores to guide learning. The method employs an information bottleneck mechanism for view reconstruction, with quality-weighted objectives applied at both feature and fusion levels. QARMVC outperformed existing methods on five benchmark datasets, particularly in scenarios with heterogeneous noise intensities.
arXiv cs.AI
·
6 hours ago
Text-to-image diffusion models released between 2022 and 2025 produce visually appealing images but perform worse as synthetic training data for computer vision classifiers despite improvements in visual quality. Classifiers trained exclusively on synthetic data from newer T2I models show consistent accuracy declines on real test data, with accuracy dropping as T2I model generations advance. The findings suggest that generative models' progress in visual fidelity does not translate to generating diverse, representative training data that matches real-world distributions, requiring reconsideration of their utility for synthetic dataset creation.
arXiv cs.AI
·
6 hours ago
Researchers introduced SEMA, a new attention mechanism that addresses the quadratic complexity of standard transformer attention and the focusing limitations of linear attention variants. SEMA combines token localization with arithmetic averaging and achieves better ImageNet-1k classification results than recent vision Mamba models at comparable parameter sizes. This approach provides a more efficient attention mechanism for computer vision tasks that maintains both local focus and global context.
arXiv cs.AI
·
6 hours ago
Scientists developed a framework that uses large language models to help researchers create 3D animations of massive climate datasets on standard computers instead of requiring specialized supercomputers and expert staff. The system processes petabyte-scale NASA climate data and generates animations in 1 minute to 2 hours by translating natural-language requests into visualization scripts without requiring visualization expertise. This reduces the time and resources needed for post-analysis visualization tasks and allows scientists to quickly share results with collaborators.
arXiv cs.AI
·
6 hours ago
Researchers introduced CXRAgent, an AI system that uses a director-coordinated agent framework to interpret chest X-rays by orchestrating multiple diagnostic tools and expert agents. The system includes three main stages: tool invocation with evidence validation, diagnostic planning that assembles task-specific expert teams, and collaborative decision-making that synthesizes insights from multiple agents. CXRAgent demonstrated strong performance across various chest X-ray interpretation tasks with improved generalization to clinical scenarios of different complexity levels.
arXiv cs.AI
·
6 hours ago
SceneBind is a multimodal representation system that combines semantic understanding with 3D spatial information across vision, audio and language, using object-centric slots alongside global embeddings. The researchers created a new binaural audio-visual dataset with spatial annotations and trained the system to align semantic and spatial signals across modalities using a lightweight token-based approach. This enables applications including cross-modal scene retrieval, object grounding, and audio-visual localization tasks.
arXiv cs.AI
·
6 hours ago
Researchers introduced Symbal, a method for detecting systematic errors in image captions generated by multimodal large language models, where the same type of error repeatedly occurs when specific visual features are present. The benchmark SymbalBench contains 1.7 million image-text pairs across 420 vision-language datasets from natural and medical domains. Symbal achieved 63.8% accuracy in identifying systematic misalignments, enabling users to audit MLLM-generated captions without access to the underlying models.
arXiv cs.AI
·
6 hours ago
Researchers released MM-IssueLoc, a benchmark containing 652 issue-PR instances across 23 languages to evaluate how well systems locate bugs in code repositories using both text and visual evidence like screenshots and error dialogs. The strongest evaluated systems achieved 38.96% file-level accuracy at rank 5 and 22.45% function-level accuracy at rank 10, showing current approaches remain far from reliable multimodal localization. The benchmark enables future research to measure whether systems actually use visual information to find issues or rely primarily on text-based approaches.
arXiv cs.AI
·
6 hours ago
Researchers developed two approaches for understanding hierarchical structures in newspaper images: a modular pipeline combining YOLO, LayoutReader, and custom algorithms, and Tiramisu, a transformer-based architecture using tiered attention mechanisms. Tiramisu performs section separation, block localization, semantic categorization, and reading order prediction through parallelized attention. A new dataset called Finlam La Liberté was released for evaluating hierarchical information retrieval in historical newspapers, with results demonstrating both methods' effectiveness in reconstructing complex newspaper hierarchies for document digitization.
arXiv cs.AI
·
6 hours ago
Researchers propose Random Logit Scaling (RLS), a defense mechanism that randomizes neural network output scores to protect deep learning models against black-box adversarial attacks. RLS significantly reduces the success rate of state-of-the-art black-box score-based attacks while maintaining model accuracy and can be implemented as a plug-and-play post-processing layer on existing models. The work also demonstrates vulnerabilities in existing non-randomized defenses by introducing an adaptive attack against the AAA defense mechanism.
arXiv cs.AI
·
6 hours ago
Researchers introduced FlashDecoder, a Transformer-based video decoder that replaces slower 3D convolutional decoders for converting latent representations to pixels in video generation. The decoder achieves 3.6x-4.7x faster decoding speeds with 11x less memory usage while matching the reconstruction quality of convolutional baselines, reaching 41.55dB PSNR at 1080p resolution on the Wan2.1 latent space. This enables real-time video generation at high resolutions with constant memory requirements regardless of video length.
arXiv cs.AI
·
6 hours ago
Researchers developed GlanceFace, a framework that infers apparent personality types from facial images alone using vision-language models and semantic-enhanced representations to capture personality-related visual cues. The system was evaluated on MBTI-based personality benchmarks and demonstrates the ability to identify relationships between facial characteristics and perceived personality traits. The approach enables embodied agents in human-robot interaction to form initial personality assessments from appearance before interaction begins, supporting adaptive interaction strategies.
