SAFE-Cascade Slashes VLM Costs by 26.9% — But Accuracy? It Stays Identical

Researchers from multiple institutions have introduced SAFE-Cascade, a cost-adaptive system that intelligently routes chart question answering queries between a cheap text-only language model and a more expensive vision-language model (VLM). The system first extracts text from chart images using OCR via Azure Document Intelligence, then feeds the text and natural language question into GPT-5-mini for a provisional answer. A trained Random Forest router—using inference-time features like OCR confidence and question complexity—decides whether to accept the text-only answer or escalate the query to Gemini-2.5-flash-image, a VLM that can visually interpret the chart.

On a held-out test set of 375 ChartQA examples, SAFE-Cascade achieved 69.1% unified accuracy while invoking the VLM only 73.1% of the time. This compares favorably to the full-VLM baseline, which hit 67.7% accuracy but required VLM calls for every query. The 1.4 percentage-point accuracy difference is statistically uncertain, meaning SAFE-Cascade effectively matches full-VLM performance while cutting VLM usage by 26.9% and estimated cost by 9.3%.

Beyond cost savings, SAFE-Cascade offers transparency and tunability. Users can inspect OCR evidence, text-only answers, routing probabilities, escalation decisions, and cost/latency estimates side by side. An adjustable escalation threshold lets system operators explore the accuracy-cost frontier—opting for cost efficiency or higher accuracy as needed. This makes SAFE-Cascade especially valuable for enterprises that process high volumes of chart questions and want to minimize VLM expense without sacrificing answer quality.