Same-class controls isolate local audio evidence in model answers
Original title: Beyond Answer Changes: Matched Same-Class Controls for Local Audio Evidence
Different-class edits change answers 49.7 to 60.9 points more often than matched same-class replacements.
The paper tests whether audio-language models change answers because of a sound event's class or because the waveform was rewritten. Matched same-class and different-class replacements share interval, level, donor fitting, and crossfade, while same-class replacements keep the correct answer. Across 600 synthetic soundscapes and three models from two architecture families, different-class edits drive far more answer changes, and same-class accuracy stays close to the original.
Unconfirmed: The three models are not named in the abstract.
- Same-class accuracy stays within 5.2 percentage points of original accuracy.
- Selective accuracy for the inserted class is 73.7 to 81.9 percent.
- Scores average eligible trials per soundscape, then weight soundscapes equally.
From the original article
1 min read at arxiv.orgFor audio-language models, an answer change after a local sound-event edit can reflect a response to the event's class, sensitivity to waveform rewriting, or both. We compare matched same-class and different-class replacements of a queried event. Both use the same interval, level estimate, donor fitting, and crossfade, while the same-class replacement preserves the correct answer.
By Yunuo Wang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.