activation patching

How audio-language models lose to text

A new paper shows audio-language models often encode the right audio answer, but text still wins the final decision.