AI TrackerFrontier desk · AEST
Back to news
normalaitech

Beyond Answer Changes: Matched Same-Class Controls for Local Audio Evidence

arXiv cs.AI · 28 Oct 2026, 03:36 AEST

A study examines how audio-language models respond to local sound-event edits. It finds that different-class replacements lead to significantly more answer changes compared to same-class replacements.

Key facts

Read aloud

A study examines how audio-language models respond to local sound-event edits. It finds that different-class replacements lead to significantly more answer changes compared to same-class replacements. For 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. Across six hundred synthetic soundscapes and three models from two architecture families, different-class replacements produce forty nine point seven to sixty point nine percentage points more answer changes than same-class replacements. Same-class accuracy remains within five point two percentage points of original accuracy. Among trials with correct original and same-class answers, soundscape-weighted selective accuracy for the inserted class is seventy three point seven to eighty one point nine percent, averaging eligible trials within each soundscape and then weighting eligible soundscapes equally.

Sources in this cluster