GRPO Tuning Increases Driving QA Accuracy Under Token Constraints

Applying GRPO to an Ornith-style VLM for multi-view driving tasks improves answer extraction from 78.25% to 90.25% within a 256-token limit. This verifiable-reward tuning helps models compress intermediate reasoning into usable answers without hitting token caps.