benchmarksfactbullishDramaSR-LRM built on large reasoning models significantly outperforms existing baselines, particularly on short utterances where acoustic biometrics are unreliableArtificial Intelligence27 Jul 2026http://arxiv.org/abs/2607.02504v1