Tag
This paper quantifies how high-performing ASR models optimize for benchmarks in ways that inflate scores without improving real-world transcription, using behavioral probes to reveal benchmark-conditioned behaviors.