用注意力图检测语音大模型幻觉，无需真实标签。
Detecting Hallucinations in SpeechLLMs at Inference Time Using Attention Maps
- 提取音频与文本注意力特征，构建轻量检测模型。
- 在语音识别任务中提升最多0.23的PR-AUC性能。
- 仅需约100个注意力头，适合实际部署。
语音大模型中的幻觉带来显著风险，而现有检测方法多依赖成本高昂或难以获取的标注输出。针对文本大模型的检测方法也难以捕捉音频特异性信号。本文研究四种基于注意力的度量：AUDIORATIO、AUDIOCONSISTENCY、AUDIOENTROPY和TEXTENTROPY，用于捕捉与幻觉相关的异常注意力模式，并在此基础上训练轻量级逻辑回归分类器，实现推理时高效检测。在自动语音识别与语音翻译任务中，对Qwen-2-Audio和Voxtral-3B的评估表明，该方法在域内数据上优于基于不确定性和先前注意力的基线，性能提升最高达+0.23 PR-AUC；同时在域外语音识别场景中仍具备良好泛化能力。进一步发现，仅使用约100个注意力头即可实现强性能，且比使用全部头更利于域外泛化。尽管效果受模型和任务影响，需针对性训练，但结果表明注意力模式是语音大模型幻觉检测的有效工具。
原文摘要 · Abstract (English)
Hallucinations in Speech Large Language Models (SpeechLLMs) pose significant risks, yet existing detection methods typically rely on gold-standard outputs that are costly or impractical to obtain. Moreover, hallucination detection methods developed for text-based LLMs do not directly capture audio-specific signals. We investigate four attention-derived metrics: AUDIORATIO, AUDIOCONSISTENCY, AUDIOENTROPY, and TEXTENTROPY, designed to capture pathological attention patterns associated with hallucination, and train lightweight logistic regression classifiers on these features for efficient inference-time detection. Across automatic speech recognition and speech-to-text translation tasks, evaluations on Qwen-2-Audio and Voxtral-3B show that our approach outperforms uncertainty-based and prior attention-based baselines on in-domain data, achieving improvements of up to +0.23 PR-AUC, and generalises to out-of-domain ASR settings. We further find that strong performance can be achieved with approximately 100 attention heads, improving out-of-domain generalisation compared to using all heads. While effectiveness is model-dependent and task-specific training is required, our results demonstrate that attention patterns provide a valuable tool for hallucination detection in SpeechLLMs.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。