Large Audio-Language Models (LALMs) excel at diverse audio tasks, yet remain prone to hallucinating sound events absent from the input. Prior work links such hallucinations to insufficient audio attention. However, we find that real and hallucinated events can receive comparable audio attention mass while attending to different audio positions. To examine whether the attended representations support the generated events, we adapt LatentLens to retrieve neighboring contextualized text representations. The retrieved contexts reveal semantic support for real events, but a semantic mismatch between hallucinated events and their attended audio representations. Directly masking the corresponding audio regions can suppress hallucinated events at the cost of real content. We are motivated by these findings to introduce an explainable and training-free detect-then-mitigate framework for effectively hallucination detection and mitigation that aggregates semantic support and applies differential audio masking to protect regions associated with events predicted to be supported. Extensive experiments on three LALMs and two datasets demonstrate that our method significantly improves hallucination detection and hallucination mitigation performance.
Figures & tables
Figure 1: Previous efforts for audio hallucination detection or mitigation.
Figure 2: A typical case with audio attention analysis. (a) Attention allocation to audio features, text prompt, and caption history across layers. The shaded band marks the Audio Attention Stage (layers 15–22). (b) Mean audio attention mass for real and hallucinated events, averaged over the Audio Attention Stage; error bars denote 95% bootstrap confidence intervals. (c) A representative response, its input Mel spectrogram with ground-truth sound event annotations, and attention maps for the real event “water flows” and the hallucinated event “goat baas” at layers 15, 19, and 21. Attention maps are cropped to 2-second windows; dashed boxes mark the most-attended audio positions, with approximate timestamps.
Figure 3: Through LatentLens, the semantics decoded from high attention frames in the Audio Attention Stage lend semantic support to real sound event tokens while withholding it from hallucinated sound event tokens.
Figure 4: Overview of our detect-then-mitigate framework. Detection (a, b): event-level attention selects audio representations, and latent semantic retrieval assesses their support for generated events. Mitigation (c): Differential audio masking targets predicted hallucination regions, protects predicted supported-event regions, and regenerates the response.
Event-Level
Sample-Level
Method
GT Hal.
Pred. Hal.
Prec.
Recall
F1
PR-AUC
GT Hal.
Pred. Hal.
Prec.
Recall
F1
PR-AUC
AudioSet Strong
Audio Flamingo 3
Random
38.7
47.9
0.3241
0.4011
0.3585
0.3535
51.0
52.0
0.3846
0.3922
0.3883
0.4329
Perplexity
88.4
0.4307
0.9842
0.5992
0.5416
94.4
0.5360
0.9922
0.6960
0.6298
RAUQ Entropy
42.2
0.5725
0.6253
0.5977
0.5814
69.0
0.6145
0.8314
0.7067
0.7037
Table 1: Hallucination detection results on AudioSet Strong and DCASE2019. Higher is better for Precision, Recall, F1, and PR-AUC. Best and second-best results within each model–level group are shown in bold and underlined, respectively.
AudioSet Strong
DCASE2019
Method
ECHO-I ↓
ECHO-S ↓
Coverage ↑
ECHO-I ↓
ECHO-S ↓
Coverage ↑
Audio Flamingo 3
Greedy
22.7
34.0
31.8
2.9
3.4
60.1
Beam
25.9
36.6
30.9
2.6
2.8
60.0
Nucleus
29.8
43.6
29.6
3.1
3.6
59.5
AAD
20.6
29.2
34.2
2.5
2.8
62.3
Table 2: Hallucination mitigation results on AudioSet Strong and DCASE2019. Metric values are percentages. Lower is better for ECHO-I and ECHO-S ( Kuan et al., 2024 ) , while higher is better for Coverage. Best and second-best results within each dataset–model group are shown in bold and underlined, respectively.
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Layer-wise attention allocation on 495 AudioCaps validation clips for each model. Attention to audio features, non-audio prompt tokens, and caption history is averaged over heads and caption-generation decisions within each clip, then equally across clips. Shaded intervals indicate the Audio Attention Stage: L18–24 for Qwen2.5-Omni-7B and L15–23 for Step-Audio 2 mini. The profiles include all caption-generation decisions, not just tokens belonging to sound events.
Dataset
Model
k
j
K
AudioSet Strong
Audio Flamingo 3
3
3
5
Qwen2.5-Omni
3
3
4
Step-Audio 2 mini
3
4
13
DCASE2019
Audio Flamingo 3
3
5
4
Qwen2.5-Omni
3
4
9
Step-Audio 2 mini
3
4
6
Appendix
Table 3: Event-level detection configurations.
Model
Retrieval
Cond.
Clip-level Hit@K
Token-level Interpretability@5
H@1
H@5
H@10
Micro
Macro
Audio Flamingo 3
Same-layer
Real
49.27
66.96
73.43
58.17
56.90
Sil.
0.89
5.96
10.84
4.69
4.86
Rand.
1.21
5.98
11.24
5.65
6.00
Shuf.
2.60
7.59
11.99
7.74
7.62
Cross-layer
Real
49.46
68.03
77.78
64.41
63.14
Appendix
Table 4: Semantic validity on AudioTime with ancestor-or-descendant matching. AF3 = Audio Flamingo 3; Qwen = Qwen2.5-Omni-7B; Step = Step-Audio 2 Mini. Real = real audio; Sil. = silence; Rand. = random retrieval; Shuf. = shuffled ground truth.
Figure 6: Precision–recall curves for hallucination detection on AudioSet Strong (left column pair) and DCASE2019 (right column pair). Rows correspond to Audio Flamingo 3, Qwen2.5-Omni, and Step-Audio 2 mini. Within each dataset, event-level and sample-level results appear on the left and right, respectively. Colors identify the five methods; markers indicate the operating points in Table 1 . Dotted horizontal lines indicate positive-class prevalence. Each panel uses the same evaluation items across methods, on the model-specific post-hoc selected cohorts underlying the table.
Model
Level
Reference F1
K±1
j±1
Layer perturbations
k=2
Audio Flamingo 3
Event
0.7466
0.7016–0.7206
0.7242–0.7244
0.7296–0.7352
0.6930
Sample
0.7897
0.7667–0.7972
0.7695-0.7695
0.7568–0.7727
0.7168
Qwen2.5-Omni
Event
0.6103
0.5880–0.5978
0.5954–0.6096
0.5885–0.5961
0.5717
Sample
0.7491
0.7367–0.7406
0.7312–0.7434
0.7159–0.7435
0.7227
Step-Audio 2 mini
Event
0.5430
0.5359–0.5486
0.5359–0.5440
0.5329–0.5412
0.4844
Sample
0.6767
0.6639–0.6682
0.6626–0.6728
0.6612–0.6680
0.6333
Appendix
Table 5: Local sensitivity of hallucination detection on AudioSet Strong. F1 is reported on a 0–1 scale. Ranges contain all available perturbed configurations for that factor, excluding the reference. Layer perturbations remove either boundary layer or shift the window by one within cached coverage. All unperturbed parameters remain fixed.
Audio Flamingo 3
Qwen2.5-Omni-7B
Step-Audio 2 mini
Event-level
Sample-level
Event-level
Sample-level
Event-level
Sample-level
k
j
F1
PR-AUC
F1
PR-AUC
F1
PR-AUC
F1
PR-AUC
F1
PR-AUC
F1
PR-AUC
1
1
0.6058
0.6185
0.6755
0.7385
0.5151
0.4643
0.6849
0.5782
0.3849
0.3845
0.5449
0.5389
2
0.6242
0.6641
0.6755
0.7922
0.5246
0.4910
0.6906
0.6026
0.3849
0.4062
0.5449
0.5639
3
0.6388
0.6881
0.6755
0.8195
0.5341
0.5138
0.6910
0.6178
0.3849
0.4196
0.5449
0.5724
4
0.6388
0.6950
0.6755
0.8266
0.5434
0.5262
0.6952
0.6327
0.3849
0.4230
0.5449
0.5673
Appendix
Table 6: Sensitivity on AudioSet Strong to audio-token budget k∈{1,2,3} and text-neighbor budget j∈{1,2,3,4,5} . The layer range is fixed. Gray cells mark the (k,j) choices achieve the results. Higher values are better.
Detection (Ours)
Mapping
Precision ↑
Recall ↑
F1 ↑
PR-AUC ↑
depth=3
0.7159
0.6517
0.6823
0.6799
ancestor
0.7888
0.6702
0.7247
0.7680
descendant
0.5924
0.7441
0.6596
0.6219
ancestor-or-descendant
0.8646
0.6570
0.7466
0.7944
exact
0.5538
0.7599
0.6407
0.5949
Appendix
Table 7: Sensitivity to AudioSet ontology mapping for AF3 (Ours). Top: event-level detection with fixed ground-truth labels and detection parameters. Bottom: mitigation evaluation of fixed responses under different mapping rules. Detection metrics are on a 0–1 scale; mitigation metrics are percentages.
ECHO-I ↓
ECHO-S ↓
Coverage ↑
Model
1
5
10
15
1
5
10
15
1
5
10
15
AudioSet Strong
Audio Flamingo 3
21.1
18.1
14.9
18.5
32.4
27.6
23.0
28.8
32.6
33.9
34.1
33.1
Qwen2.5-Omni
26.7
24.1
21.4
25.4
44.2
42.2
37.0
42.6
36.9
38.0
38.6
37.0
Step-Audio 2 mini
19.6
19.2
19.4
19.6
30.6
29.4
28.6
29.2
30.4
29.6
27.8
28.6
DCASE2019
Appendix
Table 8: Effect of the differential masking budget b on the 500-clip evaluation sets used in the main mitigation experiment. All metrics are percentages. Best and second-best results for each metric within each dataset–model group are shown in bold and underlined, respectively.
Model
Method
ECHO-I ↓
ECHO-S ↓
Coverage ↑
Audio Flamingo 3
Contrast Masking
30.1
44.2
29.7
Random Masking
26.7
39.8
31.6
Direct Masking
16.8
25.0
33.5
Differential Masking (Ours)
14.9
23.0
34.1
Qwen2.5-Omni-7B
Contrast Masking
43.0
66.4
31.4
Random Masking
38.7
57.8
32.1
Appendix
Table 9: Comparison of masking strategies on AudioSet Strong. All metrics are percentages. Bold indicates the best result within each model.
Model
Original Hal. (%)
Caption
CLAP
BLEU-4
METEOR
ROUGE-L
CIDEr
Audio Flamingo 3
24.35
Original
0.4705
0.2303
0.2526
0.4893
0.4963
After masking
0.4690
0.2785
0.2855
0.5172
0.6098
Qwen2.5-Omni-7B
41.35
Original
0.5259
0.0157
0.1409
0.1635
0.0004
After masking
0.5248
0.0168
0.1447
0.1736
0.0022
Step-Audio 2 mini
19.79
Original
0.4473
0.0474
0.1760
0.3478
0.2948
After masking
0.4626
0.0516
0.1828
0.3439
0.3333
Appendix
Table 10: Response quality before and after masking on model-specific subsets of the AudioCaps test set (we get 965/975 clips in the test set), restricted to samples whose original responses contain hallucinated sound events. CLAP compares both responses with the original audio; text metrics use all five reference descriptions. All metrics are higher-is-better.
Figure 7: Attention heatmaps, LatentLens decoding results, and hallucination mitigation for Audio Flamingo 3.
Figure 8: Attention heatmaps, LatentLens decoding results, and hallucination mitigation for Qwen2.5-Omni.
Figure 9: Attention heatmaps, LatentLens decoding results, and hallucination mitigation for Step-Audio 2 mini.
Large Audio Language Models (LALMs) achieve strong performance on audio-language tasks; however, their reliability in real-world settings remains underexplored. We introduce Audio Hallucination Attacks (AHA), an attack suite called AHA-Eval, comprising 6.5K QA pairs designed to test whether LALMs genuinely ground their responses in the audio input. AHA targets two attack surfaces: (i) query-based attacks, which exploit question structure to induce hallucinations about absent sounds, and (ii) audio-based attacks, which inject synthetic speech describing non-existent events into the audio stream. Evaluating state-of-the-art LALMs, including Audio Flamingo 3 and Gemini 3 Pro, we observe high attack success rates of 95.35% and 79.65%, respectively, revealing a reliability gap that is hidden by standard benchmark performance. To mitigate this, we propose a 120K QA post-alignment dataset, AHA-Guard, which successfully reduces attack success rates by up to 49%.
Large audio-language models (LALMs) produce fluent responses about audio but often hallucinate by making plausible yet ungrounded claims. Existing audio hallucination benchmarks mainly measure response correctness, leaving it unclear whether an LALM hallucinates or simply fails to understand the audio. We challenge correctness-based evaluation by defining two hallucination categories: (i) context, where claims are not grounded in the audio; and (ii) knowledge, where claims about audio-related topics lack support from externally verifiable facts. We introduce MISHAP-Bench, a comprehensive benchmark with 12,000 challenging open-ended question-audio pairs and a rigorous evaluation pipeline covering both categories. To evaluate open-ended responses, we propose a groundedness judge that uses reference rubrics and judge prompts guided by human annotations. We extensively evaluate ten state-of-the-art LALMs and show that hallucination remains substantial. Even a frontier model such as Gemini 3.7 Flash reaches a hallucination rate of 36.5%. We further adapt and benchmark four mitigation methods from multiple domains for LALMs. Despite some improvements, effective hallucination mitigation remains an open challenge. Finally, we call on the community to evaluate hallucination and benchmark mitigation methods with MISHAP-Bench.
Zhi Wen Soi, Giulio Segalini, Jian-Jia Chen +1
University of Neuchâtel · TU Dortmund University · RWTH Aachen University +1
Large Audio-Language Models (LALMs) excel in Audio QA but often suffer from hallucinations ungrounded in the audio. To our knowledge, we are the first to propose applying vector steering to the audio domain to mitigate this. Unlike text-based steering, our silence-anchored contrastive approach steers the model away from hallucinations by contrasting active audio against a silent baseline. Probing internal states reveals a strong correlation between specific layer representations and output correctness. Leveraging this, we introduce Layer-Weighted Vector Steering (LWVS), a training-free intervention that increases steering strength at influential layers. On the Audio Hallucination QA dataset, LWVS significantly outperforms baselines, boosting Recall on the Gemma model by 15.6% (53.4% to 69.0%). Crucially, MMAU benchmark tests confirm LWVS preserves and even enhances general audio understanding, achieving an 8% relative accuracy increase on the Qwen model (54.8% to 59.2%).
Tsung-En Lin, Kuan-Yi Lee, Hung-Yi Lee
National Taiwan University, Taipei, Taiwan · ASUS Open Cloud Infrastructure Software Center, Taipei, Taiwan