Beyond Co-existence: Measuring Attribute Binding Hallucinations in Audio-Language Models

  • Kim, Mingi; 
  • Kwon, Minchol; 
  • Ma, Minuk; 
  • Pham, Trung X.; 
  • Kim, Junyeong
Citations

WEB OF SCIENCE

0
Citations

SCOPUS

0

초록

Recent advancements in Audio-Language Models (ALMs) have enabled impressive capabilities in understanding and reasoning about complex acoustic scenes. However, these models frequently suffer from hallucinations, generating plausible but factually incorrect descriptions. Existing evaluation methods rely on discriminative, closed-ended prompts that focus on detecting the simple existence of sounds, failing to assess whether models correctly associate attributes with their sources. We define this critical failure mode as Attribute Binding Hallucination—where a model may detect both ‘man’ and ‘crying’ yet incorrectly bind them. We quantitatively expose this evaluation gap: while models achieve near-perfect accuracy on existence checks, they fail to reject mismatched attribute-object pairs with near-zero accuracy. To address this, we propose the Attribute Binding Evaluation Framework (ABEF), a methodology that measures fine-grained relational failures in open-ended responses, along with a synthetic dataset to rigorously isolate binding errors. Our experiments reveal that attribute binding hallucination is pervasive, with error rates exceeding 74%, indicating that high detection accuracy does not imply true compositional understanding. All code and datasets are available at https://github.com/1kmmk1/ALM_Hallucination.

키워드

Attribute Binding; Audio-Language-Models; Evaluation Metric; Hallucination
제목
Beyond Co-existence: Measuring Attribute Binding Hallucinations in Audio-Language Models
저자
Kim, Mingi; Kwon, Minchol; Ma, Minuk; Pham, Trung X.; Kim, Junyeong
DOI
10.1007/978-3-032-31930-2_26
발행일
2027
유형
Proceedings Paper
저널명
Lecture Notes in Computer Science
권
16825
페이지
383 ~ 396