Question guided progressive reasoning for surgical visual question localized answering

Abstract Surgical visual question localized answering (VQLA) connects language-guided scene understanding with spatial perception for context-aware surgical robotics. However, existing methods often use fixed multimodal representations, limiting adaptation of visual evidence to each question’s semantic and spatial needs. We propose SurgPR-Mamba, a question-guided progressive reasoning framework that combines Mamba-based multimodal modeling with progressive evidence refinement. Spatially aware visual features are aligned with question semantics and integrated through cross-modal Mamba interaction to establish question-relevant representations while preserving contextual dependencies. Based on this representation, visual evidence and question states are iteratively updated across multiple reasoning hops, allowing the reasoning focus to evolve from broader surgical context towards more discriminative target- and interaction-related cues. The refined evidence is subsequently used for joint answer prediction and visual localization. With 174.03M parameters, SurgPR-Mamba achieves competitive performance on EndoVis-18-VQLA and reliable generalization on EndoVis-17-VQLA. These results suggest that question-guided progressive reasoning may provide a compact and spatially grounded approach to surgical scene understanding for context-aware robotic assistance.