Understanding Sounds, Missing the Questions: The Challenge of Object Hallucination in Large Audio-Language Models

TL;DR

This work investigates object hallucination in large audio-language models (LALMs) by introducing discriminative and generative evaluation tasks grounded in AudioCaps and CHIME-6. It demonstrates that while LALMs can match specialized audio captioning models in understanding audio content, they struggle with discriminative questions and exhibit prominent object hallucination; prompt engineering can mitigate some of these failures, though a sizable gap remains compared to cascade approaches. The study introduces rigorous metrics (ECHO and Cover) and sampling strategies to quantify hallucination, and shows that LALMs’ reliability can be improved via carefully designed prompts. Overall, the paper highlights a critical reliability gap in LALMs and provides concrete evaluation protocols and mitigation strategies to bridge it in practice.

Abstract

Large audio-language models (LALMs) enhance traditional large language models by integrating audio perception capabilities, allowing them to tackle audio-related tasks. Previous research has primarily focused on assessing the performance of LALMs across various tasks, yet overlooking their reliability, particularly concerning issues like object hallucination. In our study, we introduce methods to assess the extent of object hallucination of publicly available LALMs. Our findings reveal that LALMs are comparable to specialized audio captioning models in their understanding of audio content, but struggle to answer discriminative questions, specifically those requiring the identification of the presence of particular object sounds within an audio clip. This limitation highlights a critical weakness in current LALMs: their inadequate understanding of discriminative queries. Moreover, we explore the potential of prompt engineering to enhance LALMs' performance on discriminative questions.

Figures (1)

• Figure 1: Demonstration of our evaluation pipeline.