SAFE-Cascade: Cost-Adaptive Vision-Language Routing for Chart Question Answering

A new demonstration system called SAFE-Cascade shows how chart question answering can be made cheaper by routing simple queries to a text-only model and reserving vision-language models for cases that actually require visual reasoning, according to research posted on arXiv [1]. The system, described in a paper submitted June 17, extracts text from chart images using optical character recognition and obtains a provisional answer from a text-only language model. A learned router then decides whether to accept that answer or escalate the query to a vision-language model, or VLM [1]. The implementation uses Azure Document Intelligence for OCR, gpt-5-mini as the text-only model, and gemini-2.5-flash-image as the VLM, with a Random Forest router trained on inference-time features [1]. On a held-out ChartQA test split of 375 examples drawn from a 2,500-example experiment, SAFE-Cascade achieved 69.1% unified accuracy while invoking the VLM on 73.1% of queries [1]. The full-VLM baseline reached 67.7% accuracy with 100% VLM invocation. The authors note that the 1.4 percentage-point difference is statistically uncertain and interpret the result as matching full-VLM performance while cutting VLM calls by 26.9% and estimated cost by 9.3% [1]. The demonstration interface exposes the entire decision process. Users can inspect OCR evidence, the text-only answer, routing probability, escalation decision, final answer, estimated cost, and estimated latency side by side [1]. An adjustable escalation threshold lets users explore the accuracy-cost frontier directly [1]. Vision-language models combine visual encoders with large language models to process images and text jointly. They belong to a broader lineage of neural network architectures that gained traction through convolutional neural networks for vision tasks and transformer-based attention mechanisms that now underpin most large language models [3]. The SAFE-Cascade approach reflects a growing interest in selective modality routing — only paying the computational cost of a full multimodal model when simpler text-based reasoning proves insufficient [1]. The paper appears on arXiv, a preprint repository that has become a central distribution channel for machine learning research. Platforms such as Hugging Face have built infrastructure to link arXiv papers with community-contributed demos, models, and datasets, allowing readers to interact with systems like SAFE-Cascade directly in a browser [8][9].