Uncertainty Estimation Benchmarking for Audio Question Answering

A comparative study of five uncertainty estimation methods across four open-weight audio models shows that first-token probability achieves a mean AUROC of 0.740 in multiple-choice tasks. This approach outperforms ten-sample discrete semantic entropy while requiring zero additional model calls.