Compared with typical vision–language tasks, document parsing places stronger demands on fine-grained visual perception. Existing vision–language model (VLM)–based parsing approaches rely on globally compressed visual tokens, where fine-grained details are entangled within a single representation and repeatedly accessed during decoding. However, we observe that the visual evidence for each prediction is typically localized and conditioned on the current decoding state, whereas such representations must be accessed in full at every decoding step, resulting in inefficient computation. To address this mismatch, we formulate perception as state-conditioned visual evidence retrieval (SCVER) during autoregressive decoding. The model operates on a compact global representation for coarse structure and retrieves a small set of relevant high-resolution regions conditioned on the current token state. This coarse-to-fine design enables on-demand access to fine-grained visual cues, relieving globally shared representations from encoding all fine-grained details. We further find that learning such state-conditioned retrieval in VLMs is challenging and unstable. To stabilize this process, we introduce a Spatially-Guided Learning Objective (SGLO) to guide the retrieval process. Experiments on document parsing benchmarks show that SCVER improves robustness under reduced input resolution and achieves a better accuracy–efficiency trade-off, demonstrating the effectiveness of on-demand visual evidence retrieval for fine-grained perception.

Three-dimensional Computed Tomography (3D CT) is a cornerstone of precision medicine. Most AI diagnostic models analyze large num bers of CTslices uniformly, treating all slices as equally important. While this has partly accel erated radiologists’workflows, it overlooks that clinically relevant information is often sparsely distributed throughout a volume. Without tar geted or weighted processing, fine-grained cues may be missed and substantial computation wasted on diagnostically uninformative slices. Wepropose aradiologist-simulating framework for selective and efficient 3D CT interpreta tion. Evaluated on a 3D CT dataset covering eight thoracic lesion types, it was compared with state-of-the-art multimodal large language models such as GPT-4o and supervised visual backbones including ViT and ResNet-50. Us ing accuracy, F1-score, AUC, and blind radiolo gist assessment, Screen-CLIP achieved an AUC of 0.87 and F1-score of 0.82, surpassing ViT Base (AUC: 0.84). For report generation, our method outperformed M3D across all metrics, reaching a BLEU-Avg of 29.03, and achieved the highest average Doctors’ Score (6.16/10) in a preliminary human evaluation.