Detecting Hallucinations in SpeechLLMs at Inference Time Using Attention Maps

• URL: http://arxiv.org/abs/2604.19565v1
• Date: Tue, 21 Apr 2026 15:18:10 GMT
• Title: Detecting Hallucinations in SpeechLLMs at Inference Time Using Attention Maps
• Abstract summary: Hallucinations in Speech Large Language Models (SpeechLLMs) pose significant risks, yet existing detection methods rely on gold-standard outputs that are costly or impractical to obtain.<n>We investigate four attention-derived metrics: AUDIORATIO, AUDIOCONSISTENCY, AUDIOENTROPY, and TEXTENTROPY, designed to capture pathological attention patterns associated with hallucination.<n>Across automatic speech recognition and speech-to-text translation tasks, our approach outperforms uncertainty-based and prior attention-based baselines.
• Score: 1.8825446478246406
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: 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.

Related papers

• Toward Generalizable Cognitive Impairment Detection with Speech-Based Multimodal Large Language Models [6.567438338350837]
  Speech signals encode both linguistic and acoustic markers associated with cognitive decline.<n>Recent advances in large language models (LLMs) strengthen the potential of speech-based assessment.<n>We propose a multimodal CI detection framework based on open-source LLMs that integrates speech audio and corresponding transcripts.<n>Our work establishes a new state-of-the-art for CI identification, with the proposed method demonstrating superior cross-dataset generalization.
  arXiv  Detail & Related papers  (2026-07-23T16:43:12Z)
• Scalable Token-Level Hallucination Detection in Large Language Models [63.3426544914783]
  Internal hallucinations are difficult to detect in reasoning-intensive tasks.<n>We propose TokenHD, a holistic pipeline for training token-level hallucination detectors.
  arXiv  Detail & Related papers  (2026-05-12T16:47:40Z)
• RAGognizer: Hallucination-Aware Fine-Tuning via Detection Head Integration [0.2696472814555309]
  We introduce RAGognizer, a hallucination-aware fine-tuning approach that integrates a detection head into an Large Language Models.<n> RAGognizer achieves state-of-the-art token-level hallucination detection while substantially reducing hallucination rates during generation.
  arXiv  Detail & Related papers  (2026-04-17T11:07:32Z)
• Detecting Contextual Hallucinations in LLMs with Frequency-Aware Attention [27.49425252327799]
  We introduce a frequency-aware perspective on attention by analyzing its variation during generation.<n>We develop a lightweight hallucination detector using high-frequency attention features.
  arXiv  Detail & Related papers  (2026-02-20T11:18:45Z)
• Measuring the Impact of Lexical Training Data Coverage on Hallucination Detection in Large Language Models [26.89705770151822]
  Hallucination in large language models (LLMs) is a fundamental challenge, particularly in open-domain question answering.<n>Prior work attempts to detect hallucination with model-internal signals such as token-level entropy or generation consistency.<n>We investigate whether data coverage itself can serve as a detection signal.
  arXiv  Detail & Related papers  (2025-11-22T06:59:55Z)
• Temporal-Aware Iterative Speech Model for Dementia Detection [0.0]
  Current methods for automated dementia detection using speech rely on static, time-agnostic features or aggregated linguistic content.<n>We introduce TAI-Speech, a Temporal Aware Iterative framework that dynamically models spontaneous speech for dementia detection.<n>Our work provides a more flexible and robust solution for automated cognitive assessment, operating directly on the dynamics of raw audio.
  arXiv  Detail & Related papers  (2025-09-26T01:56:07Z)
• Zero-Shot Cognitive Impairment Detection from Speech Using AudioLLM [9.84961079811343]
  Speech has gained attention as a non-invasive and easily collectible biomarker for assessing cognitive decline.<n>Traditional cognitive impairment detection methods rely on supervised models trained on acoustic and linguistic features extracted from speech.<n>We propose the first zero-shot speech-based CI detection method using the Qwen2- Audio AudioLLM, a model capable of processing both audio and text inputs.
  arXiv  Detail & Related papers  (2025-06-20T01:28:43Z)
• Learning Auxiliary Tasks Improves Reference-Free Hallucination Detection in Open-Domain Long-Form Generation [78.78421340836915]
  We systematically investigate reference-free hallucination detection in open-domain long-form responses.<n>Our findings reveal that internal states are insufficient for reliably distinguishing between factual and hallucinated content.<n>We introduce a new paradigm, named RATE-FT, that augments fine-tuning with an auxiliary task for the model to jointly learn with the main task of hallucination detection.
  arXiv  Detail & Related papers  (2025-05-18T07:10:03Z)
• $C^2$AV-TSE: Context and Confidence-aware Audio Visual Target Speaker Extraction [80.57232374640911]
  We propose a model-agnostic strategy called the Mask-And-Recover (MAR)<n>MAR integrates both inter- and intra-modality contextual correlations to enable global inference within extraction modules.<n>To better target challenging parts within each sample, we introduce a Fine-grained Confidence Score (FCS) model.
  arXiv  Detail & Related papers  (2025-04-01T13:01:30Z)
• Where are we in audio deepfake detection? A systematic analysis over generative and detection models [59.09338266364506]
  SONAR is a synthetic AI-Audio Detection Framework and Benchmark.<n>It provides a comprehensive evaluation for distinguishing cutting-edge AI-synthesized auditory content.<n>It is the first framework to uniformly benchmark AI-audio detection across both traditional and foundation model-based detection systems.
  arXiv  Detail & Related papers  (2024-10-06T01:03:42Z)
• Training-Free Deepfake Voice Recognition by Leveraging Large-Scale Pre-Trained Models [52.04189118767758]
  Generalization is a main issue for current audio deepfake detectors. In this paper we study the potential of large-scale pre-trained models for audio deepfake detection.
  arXiv  Detail & Related papers  (2024-05-03T15:27:11Z)
• Enhancing Uncertainty-Based Hallucination Detection with Stronger Focus [99.33091772494751]
  Large Language Models (LLMs) have gained significant popularity for their impressive performance across diverse fields. LLMs are prone to hallucinate untruthful or nonsensical outputs that fail to meet user expectations. We propose a novel reference-free, uncertainty-based method for detecting hallucinations in LLMs.
  arXiv  Detail & Related papers  (2023-11-22T08:39:17Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.