Audio-language embedding models fail to encode negated sound concepts

A study reveals that audio-language embedding models like CLAP struggle with negation, mapping affirmative and negated captions to nearly identical representations. To expose this limitation, the authors introduce NegEval-Audio, a framework converting datasets into Retrieval-Neg and Multiple-Choice Negation tasks.

• Performance on AudioCaps and Clotho degrades sharply under negation, with MCQ accuracy falling below chance.
• The failure persists even for recent multimodal LLM-based embedding models.
• A training-free steering method improves MCQ-Neg but yields marginal gains for Retrieval-Neg.

The results indicate that affirmation bias is a fundamental flaw in representation geometry, necessitating explicit negation-aware training objectives.