Multimodal large language models (MLLMs) have achieved impressive performance in understanding and describing visual content, setting new state-of-the-art results on a variety of visual question answering (VQA) benchmarks. However, during decoding, these models often fail to attend to fine-grained visual details in the input image. Our analysis of intermediate attention layers reveals that MLLMs are not inherently incapable of perceiving target objects; rather, attention to visual details becomes diluted in deeper layers due to the dominance of language priors. To address this limitation, we propose a plug-and-play Attention Re-Alignment module (ARA) that enhances suppressed visual grounding. ARA conducts a layer-wise analysis of the relative attention distribution of image-centric attention heads. It incorporates a confidence-aware layer selection mechanism based on attention peak and entropy, enabling the dynamic aggregation of attention maps from the most informative layers. These aggregated maps are subsequently leveraged to guide the generation of semantic masks, enabling the model to emphasize salient visual regions while suppressing irrelevant or noisy content. ARA can be seamlessly integrated into existing MLLMs and demonstrates consistent improvements across multiple VQA benchmarks, validating its effectiveness in enhancing visual detail sensitivity.
The automatic generation of stable robotic grasping postures is crucial for the application of computer vision algorithms in real-world settings. This task becomes especially challenging in complex environments, where accurately identifying the geometric shapes and spatial relationships between objects is essential. To enhance the capture of object pose information in 3D visual scenes, we propose a planar robotic grasping detection algorithm named SU-Grasp, which simultaneously focuses on local regions and long-distance relationships. Built upon a U-shaped network, SU-Grasp introduces a novel dual-stream encoding strategy using the Swin Transformer combined with spatial semantic enhancement. Compared to existing baseline methods, our algorithm achieves superior performance across public datasets, simulation tests, and real-world scenarios, highlighting its robust understanding of complex spatial environments.
Automated radiology report generation has become a prominent research topic in medical multimodal learning. However, most existing approaches primarily focus on single-image interpretation and rarely address the task of tracking disease progression across longitudinal chest X-rays. This task presents two major challenges: accurately localizing pathological changes between temporally paired images, and effectively translating visual difference representations into clinically meaningful textual descriptions. To address these challenges, we propose MRID (Modeling Radiological Image Differences for Disease Progression Reasoning), a multi-task self-supervised framework that follows a pretraining–finetuning paradigm. MRID leverages multiple complementary self-supervised objectives to jointly achieve (1) intra-modal spatial alignment of organs and pathological regions across image pairs, and (2) cross-modal semantic alignment between visual difference representations and radiology report embeddings. Furthermore, we introduce a simple yet effective data augmentation strategy to alleviate the imbalance of disease progression categories. Extensive experiments conducted on the Longitudinal-MIMIC and MS-CXR-T datasets demonstrate that MRID effectively captures fine-grained disease progression patterns. In addition, the proposed framework achieves competitive performance on single-image radiology report generation, further highlighting its strong capability in modeling chest X-ray semantics.