Multimodal Large Language Models for Visual Attribute Inference in iRAP Road Attribute Coding

Road safety assessment is essential for reducing traffic fatalities, with road infrastructure contributing to a substantial proportion of crashes worldwide. International frameworks such as the International Road Assessment Program (iRAP) define standardized attributes for infrastructure auditing; however, many of these attributes remain challenging to automate using imagery alone. This study evaluates V-RoAst (visual question answering for road assessment), a public dataset of road images that are annotated with iRAP-style attributes, using state-of-the-art multimodal large language models (MLLMs), specifically Gemini 2.0 and Gemini 2.5. The analysis focuses on how prompt design influences the accuracy and stability of single image iRAP inference. A token-efficient reduced prompt is developed that preserves the iRAP schema while removing single-class constants, hard-coded administrative fields, and derived or non-visual codes, retaining only visually interpretable attributes. Performance is compared with the original full multi-attribute prompt and single attribute prompts using a fixed evaluation protocol incorporating majority voting, bootstrap 95% confidence intervals, and per-code sample-size checks. Results indicate only minor performance differences between Gemini 2.0 and Gemini 2.5, while prompt optimization produces the most consistent gains, improving macro-F1 scores and tightening confidence intervals for visually grounded attributes such as roadside severity, intersection channelization, and service-road presence. Token analysis shows an approximate 30% reduction in prompt length, reducing computational cost and truncation risk. Overall, the findings demonstrate that prompt scope has a greater impact than model version in image-only iRAP coding, offering practical guidance for scalable infrastructure assessment.