Reducing Object Hallucination in Large Audio-Language Models via Audio-Aware Decoding
Introduced Audio-Aware Decoding (AAD), improving LALMs' F1 score by 0.046 to 0.428 on object hallucination datasets.
Key Findings
Methodology
Audio-Aware Decoding (AAD) is a lightweight inference strategy using contrastive decoding to compare token prediction logits with and without audio context. AAD promotes tokens whose probability increases when audio is present, reducing hallucination.
Key Results
- AAD improved F1 scores by 0.046 to 0.428 on object hallucination datasets, significantly reducing hallucinations.
- On Clotho-AQA dataset, AAD improved accuracy by 5.4% to 10.3%, demonstrating effectiveness in general audio QA tasks.
- Extensive ablation studies validated the effectiveness of each AAD component.
Significance
This research introduces AAD to address hallucination issues in LALMs, enhancing performance on object hallucination and general audio QA tasks, with broad academic and industrial applications.
Technical Contribution
AAD significantly reduces hallucinations in LALMs using contrastive decoding, offering a more robust solution compared to existing prompt engineering methods, validated across multiple models and datasets.
Novelty
AAD is the first to apply contrastive decoding to LALMs, differing from previous prompt engineering methods by reducing hallucinations through audio context comparison.
Limitations
- AAD requires two model forward passes, increasing computational overhead, but can be mitigated by parallel processing.
- Current study is limited to yes/no questions, future work should expand to more complex QA scenarios.
Future Work
Future work can explore different blank audio generation strategies and applications in other audio QA tasks like MMAU or SAKURA.
AI Executive Summary
Large Audio-Language Models (LALMs) excel in audio comprehension tasks but suffer from hallucination issues, where models may incorrectly identify objects not present in the audio. Existing prompt engineering methods can reduce hallucinations to some extent but are unstable and heavily reliant on prompt design.
To tackle this challenge, researchers introduced Audio-Aware Decoding (AAD), a lightweight inference strategy using contrastive decoding to compare token prediction logits with and without audio context. AAD promotes tokens whose probability increases when audio is present, effectively reducing hallucination.
Experimental results show that AAD significantly improves F1 scores on object hallucination datasets and enhances accuracy on Clotho-AQA datasets, demonstrating its effectiveness in general audio QA tasks. This method not only reduces hallucinations but also boosts overall model performance.
However, AAD requires two model forward passes, increasing computational overhead, but can be mitigated by parallel processing. Future work can explore different blank audio generation strategies and applications in other audio QA tasks.
Overall, Audio-Aware Decoding provides a robust solution to the hallucination problem in LALMs, with broad application potential.
Deep Dive
Abstract
Large Audio-Language Models (LALMs) can take audio and text as the inputs and answer questions about the audio. While prior LALMs have shown strong performance on standard benchmarks, there has been alarming evidence that LALMs can hallucinate what is presented in the audio. To mitigate the hallucination of LALMs, we introduce Audio-Aware Decoding (AAD), a lightweight inference-time strategy that uses contrastive decoding to compare the token prediction logits with and without the audio context. By contrastive decoding, AAD promotes the tokens whose probability increases when the audio is present. We conduct our experiment on object hallucination datasets with three LALMs and show that AAD improves the F1 score by 0.046 to 0.428. We also show that AAD can improve the accuracy on general audio QA datasets like Clotho-AQA by 5.4% to 10.3%. We conduct thorough ablation studies to understand the effectiveness of each component in AAD.