Radiology report generation requires precise alignment between medical imaging findings and clinically coherent textual descriptions. While current methods predominantly rely on either large vision-language models (LVLMs) for visual grounding or large language models (LLMs) for medical narrative generation, they often fail to effectively integrate multimodal clinical evidence with domain-specific knowledge. This paper proposes a novel multimodal dual-path framework that synergistically combines LVLMs and LLMs to address these limitations. Our approach establishes a dynamic fusion between LVLMs’ visual-semantic grounding capabilities and LLMs’ clinical knowledge reasoning. Specifically, we employ a structured prompting strategy that models the report generation task into three clinically meaningful sections and introduces fine-grained multi-label classification prompts to guide the models, enabling more accurate and comprehensive clinical report generation. Experiments on the public MIMIC-CXR benchmark demonstrate our framework’s superiority over state-of-the-art methods.

Vision-language models (VLMs) often answer visual questions using learned language and category priors rather than grounding their predictions in the image itself. Counterfactual images provide a natural diagnostic setting for this failure mode: when visible evidence contradicts what is usually true, a grounded model should answer from the pixels, while a prior-following model will produce a canonical but visually incorrect response. However, existing counterfactual benchmarks mainly ask whether such prior-following behavior exists. In this paper, we ask a further question motivated by the rise of tool-augmented and agentic vision systems: can additional visual evidence views help VLMs reason against their priors? We introduce PriVE-Bench, a Prior-vs-Visual Evidence Benchmark that uses paired original and counterfactual images to distinguish visually grounded answers from prior-consistent errors. We further introduce PriVE-Tools, a controlled agentic-vision-inspired extension that evaluates whether tool-derived visual evidence—including bounding boxes, crops, zoom panels, and contours—improves grounding under the same counterfactual conflicts. Across open- and closed-source VLMs, we compare raw, paired-image, and tool-conditioned inputs using accuracy, prior-following error rate, and other-response rate. Our results show that visual evidence tools can help in some settings, especially when models can use localized evidence effectively, but they are not a universal remedy: several models continue to follow language and category priors even when relevant visual evidence is explicitly provided. The benchmark and codebase can be accessed at https://github.com/JingyuSunUOE/counterfactual-vlm-benchmark