This research proposes a systematic framework for quantifying hallucination rates in Vision-Language Models (VLMs) when deployed for aerial navigation tasks. Through controlled ablation studies across visual, linguistic, and navigational components, the paper establishes metrics to measure and mitigate hallucinations in safety-critical drone operations.
Key findings
Develops metrics for quantifying hallucination rates specific to aerial navigation contexts.
Designs controlled ablation studies to isolate factors contributing to hallucination formation.
Establishes benchmark evaluations across multiple aerial navigation datasets.
Analyzes the relationship between hallucination rates and navigation performance.
Proposes mitigation strategies based on empirical findings.
Limitations & open questions
The analysis remains limited to image captioning tasks and does not address sequential decision-making requirements of navigation.