Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

Reducing Object Hallucination in Large Audio-Language Models via Audio-Aware Decoding

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Audio-Aware Decoding reduces object hallucination in large audio-language models by contrasting real audio against silence, improving F1 by up to 0.428 on hallucination tests and accuracy by up to 10.3% on general audio QA.

desk verdict A solid, honest port of Context-Aware Decoding to audio LALMs: the empirical gains are real and consistent, but the blank-audio baseline and lack of variance reporting leave the mechanism under-validated. read the letter →

arxiv 2506.07233 v2 pith:ZRIENHZE submitted 2025-06-08 eess.AS cs.CL

Tzu-wen Hsu , Ke-Han Lu , Cheng-Han Chiang , Hung-yi Lee This is my paper
classification eess.AScs.CL
keywords largeaudio-languagemodelsobjecthallucinationcontrastivedecodingaudio-awareinference-timeaudioquestionansweringmitigationyes/noQA
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Large audio-language models sometimes answer "yes" when asked whether an object appears in a sound that does not contain it. This paper introduces Audio-Aware Decoding (AAD), an inference-time strategy that contrasts the model's token predictions with the real audio against its predictions with a silent, zero-filled audio clip of the same length, and amplifies only the tokens that become more likely when the audio is actually present. The authors claim this reduces object hallucination across three LALMs and sampling regimes, raising F1 by 0.046 to 0.428 on the hallucination benchmark while also improving accuracy by 5.4% to 10.3% on the general audio-QA dataset Clotho-AQA. If right, it gives a lightweight, model-agnostic complement to prompt engineering, which the paper shows is fragile when the absent objects are chosen adversarially.

What carries the argument

The load-bearing mechanism is the contrastive decoding identity in Equation 2: subtract the logits produced with a blank audio clip from the logits produced with the real clip, then apply a softmax with a scaling factor $\alpha$. The blank $A_{\mathrm{blank}}$ is a same-length all-zero copy of the original audio, and a fixed prefix prompt, "Focus on the given audio and answer the following question," is prepended to the question. This construction isolates the per-token logit shift caused by the audio content, so a token that only becomes likely because the sound is present gets up-weighted, while a token produced from language priors alone is down-weighted. The hyperparameter $\alpha$ sets the strength of the contrast, and the ablation study finds that $\alpha \approx 1.0$ best balances rejecting hallucinated objects with still saying "yes" when the sound is genuinely there.

What would settle it

Replace the zero-filled blank audio in AAD with a low-amplitude noise clip or a constant nonzero signal of the same length and rerun the object hallucination benchmark: if the F1 gains shrink or disappear, the method's effect depends on the specific blank-audio artifact, and if the gains persist, the all-zero baseline is not essential to the result.

Watch

Extended reading notes

Core claim

The paper's central claim is that a contrastive decoding rule, $p_{\mathrm{AAD}} = \mathrm{softmax}((1+\alpha)\, \mathrm{logit}_{\mathrm{with\text{-}audio}} - \alpha\, \mathrm{logit}_{\mathrm{without\text{-}audio}})$, makes large audio-language models answer more faithfully to the audio. Here the "without-audio" condition is the same question with an all-zero audio of identical length, and $\alpha$ controls how strongly audio-grounded evidence overrides the model's prior. Tokens whose probability rises when real audio is present are promoted, while tokens the model would emit regardless of the audio are suppressed, which counteracts the strong default bias toward answering "yes." Tested on SALMONN-7B, SALMONN-13B, and Qwen2-Audio-7B-Instruct, the authors report F1 gains of 0.046 to 0.428 on the object hallucination benchmark and accuracy gains of 5.4% to 10.3% on Clotho-AQA, with best results near $\alpha = 1$.

Load-bearing premise

The load-bearing premise is that a same-length zero-filled audio clip is a neutral and faithful "no audio" condition, so the logit difference between real audio and that blank isolates audio-grounded evidence rather than artifacts of feeding the encoder an out-of-distribution input.

Editorial extensions

If this is right

  • AAD reduces object hallucination across LALMs with different sizes and architectures, so the benefit is not tied to one model's training recipe.
  • Under adversarial and popular object sampling, where prompt-only baselines often fail or hurt, AAD still improves F1, making it more robust than prompt engineering.
  • AAD improves accuracy on Clotho-AQA as well, indicating that suppressing hallucinated "yes" answers does not damage general audio question answering.
  • The optimal contrast strength sits around $\alpha = 1$, where the model's "yes" rate falls from roughly 80-90% toward the balanced 50% of the benchmark.
  • Running the two forward passes in parallel keeps the added latency near 19%, so the method is practical for inference-time use.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same contrastive recipe could transfer to other modalities, such as video or images, whenever a blank input can be constructed that the encoder treats as content-free; that transfer is not tested in the paper.
  • The results suggest a large portion of LALM object hallucination is a prior "yes" bias rather than a failure of audio encoding, pointing toward training-time calibration as a complementary fix.
  • Because the paper only evaluates balanced yes/no benchmarks, AAD's effect on imbalanced real-world queries is unknown; pushing the yes rate toward 50% could be harmful where "yes" is rare.
  • An immediate experiment is to replace the zero-filled blank with low-level noise or a constant-offset audio; if the gains persist, the zero baseline is incidental, and if they vanish, part of the effect is an encoder artifact.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper proposes Audio-Aware Decoding (AAD), an inference-time contrastive decoding method for large audio-language models (LALMs). At each generation step, AAD computes logits both with the real input audio and with a zero-filled 'blank' audio of the same length, and reweights the token distribution as p_AAD = softmax((1+alpha) logit_with_audio - alpha logit_without_audio) (Eq. 2). The method is intended to promote tokens whose likelihood increases in the presence of audio, thereby reducing object hallucination. The authors evaluate AAD on the object hallucination benchmark of [23] with SALMONN-7B, SALMONN-13B, and Qwen2-Audio-7B-Instruct, across random, adversarial, and popular sampling of absent objects, and on the Clotho-AQA yes/no QA subset. They report F1 improvements of 0.046 to 0.428 on the hallucination benchmark and accuracy gains of 5.4% to 10.3% on Clotho-AQA, along with ablations on the hyperparameter alpha and the prefix prompt.

Significance. If the empirical results hold, AAD is an appealingly simple and training-free intervention that can be applied to any LALM at decoding time to reduce hallucination, with no modifications to model weights. The paper's main strengths are its breadth across three different LALM architectures and three sampling regimes, the consistency of the gains (especially in the random and popular sampling conditions), and the released code on GitHub. The ablations showing that AAD remains effective with a suboptimal prompt and without a prompt are useful for understanding the method's robustness. However, the central assumption that a zero-filled audio clip serves as a neutral 'no-audio' condition is not validated; this assumption underpins the mechanistic claim that AAD amplifies audio-grounded evidence. The lack of statistical significance testing and the test-set-based selection of alpha further temper the strength of the reported gains.

major comments (5)
  1. [Section II-B, Eq. (2); Section VII; Table II] The method's central assumption is that a zero-filled audio clip A_blank is a neutral and faithful 'no-audio' condition. The paper provides no evidence that the audio encoders of SALMONN, Qwen2-Audio, or similar LALMs treat an all-zero waveform as the absence of audio; the encoders may instead process it as an out-of-distribution input with normalization, positional embeddings, and audio-specific input projections. In that case, logit_without-audio may carry encoder artifacts rather than the model's no-audio prior, and the contrast in Eq. (2) could amount to subtracting a constant bias that shifts the decision threshold toward 'no'. The authors themselves list alternative blank-audio constructions only as future work in Section VII, so this load-bearing baseline is untested. I recommend adding experiments with at least one alternative blank-audio construction (e.g., a non-zero constant, low-level noise, or an explicit 'no audio' special token) and, if they yield different results, discussing the implications for the mechanism claim.
  2. [Section III-C and Table I] The paper reports no error bars, confidence intervals, or statistical significance tests for any of the F1 or accuracy differences. Some of the claimed improvements are small (e.g., +0.046 F1 for SALMONN-13B under popular sampling with alpha=0.5, going from 0.393 to 0.439) and could be within sampling variability even though the benchmark contains roughly 30,000 samples. The authors should report standard deviations or bootstrap confidence intervals and, ideally, a paired significance test (e.g., McNemar's test on the yes/no answer decisions) for at least the headline comparisons (Default vs. AAD and Prompt-only vs. AAD) on each dataset and sampling strategy.
  3. [Section V-A, Table II; Section IV] The hyperparameter alpha is selected based on performance on the same benchmark used to report the main results. The paper states in Section V-A that 'the sweet spot appears around alpha=1.0' after evaluating alpha in {0.5, 1.0, 1.5, 2.0} on the test benchmarks, and Table I then highlights AAD with alpha=1.0 (and alpha=0.5 in the abstract). This creates a risk of optimistic bias from test-set tuning. Although Table II shows that a range of alpha values improves over the default, which partially mitigates the concern, the authors should either fix alpha a priori, use a validation split, or present the results as a sensitivity analysis rather than as a tuned best configuration. The abstract's F1 range also corresponds to alpha=0.5 only, which should be stated explicitly.
  4. [Abstract and Section IV] The abstract's reported improvements do not match the tables in a transparent way. The range '0.046 to 0.428' corresponds to alpha=0.5 in Table I (e.g., SALMONN-13B popular sampling: 0.393 to 0.439, and SALMONN-7B random sampling: 0.233 to 0.661), while alpha=1.0 yields a range of roughly 0.076 to 0.504. Similarly, the conclusion's statement 'with an average boost of F1 score from 0.118 to 0.281' does not correspond to any obvious aggregate in Table I. The authors should either state the alpha value used for the summary numbers and provide a clear aggregation rule, or report the ranges separately for each alpha setting.
  5. [Section IV-B and Table I] The paper focuses on F1 for the hallucination benchmark, but accuracy is also reported and sometimes decreases with AAD. For example, SALMONN-7B under adversarial sampling has accuracy 0.509 with the default setting and 0.482 with AAD alpha=1.0, while F1 improves from 0.177 to 0.456. This indicates a trade-off between avoiding false 'yes' answers and preserving correct 'yes' answers. The paper should explicitly discuss this trade-off, especially because the abstract claims accuracy improvements only for Clotho-AQA but not for the hallucination benchmark. Practitioners need to know that AAD may sacrifice overall correctness in adversarial settings.
minor comments (6)
  1. [Abstract and Section II-B] The phrase 'A_blank is an audio without any sound' is imprecise; Section II-B immediately clarifies that it is implemented as an all-zero copy of A. The paper could state at the outset that the 'no-audio' condition is the all-zero waveform, rather than implying a semantically neutral silence representation.
  2. [Section III-C] The claim that 'random guessing on this dataset will result in an accuracy of 0.500 and an F1 of 0.500' is only true for a specific random policy (uniform random yes/no). A model that always answers 'no' would achieve accuracy 0.500 but F1 = 0.667 under the 'no-as-positive' convention. Please clarify the assumed random policy.
  3. [Table III] The delta columns labeled '∆Acc(%)/∆F1(%)' use percentage-point differences (e.g., -4.9 for 0.656 vs. 0.705) but the '%' sign suggests relative percent change. Please label as percentage points or compute relative changes consistently.
  4. [Section V-C] The statement that 'the performance of AAD is always better than the default and the prompt engineering only baseline' is too strong: in Table IV, for the 'Listen' prompt with adversarial sampling, AAD alpha=1.0 has accuracy 0.457, which is lower than the default accuracy of 0.487. The claim should be qualified to F1 or to 'on the primary F1 metric'.
  5. [Section VII] There is a typo in the limitations paragraph: 'we need to forward the through the LALM twice' should be 'forward the input through the LALM twice'.
  6. [References] Reference [23] is the authors' own prior work and is used both as the source of the benchmark and the prefix prompt; this is legitimate but should be acknowledged explicitly as a self-citation in the experimental setup or related work, since the prompt-only baseline is therefore partially tuned by the same group on the same benchmark.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: AAD empirically applies a standard contrastive-decoding formula; the only self-citation is a non-load-bearing benchmark/prompt source.

full rationale

The paper's central claim is empirical: applying Equation 2 with a zero-filled blank audio improves F1/accuracy on two benchmarks. No prediction is derived from an equation that already contains the target result. Equation 2 is the standard contrastive-decoding formula, explicitly attributed to [30], [33], [37], [38]; AAD is an application, not a derivation. The only self-citation is [23], which supplies the hallucination benchmark and the prefix prompt; this is not load-bearing because the same gains appear on the independent Clotho-AQA benchmark and with a different prompt ('Listen') and with no prompt (Tables III-IV). The unvalidated assumption that an all-zero audio is a neutral no-audio condition (Section II-B) is a correctness/validity concern, not circularity, and the paper itself flags alternative blank-audio constructions as future work in Section VII. Alpha is varied over 0.5-2.0 and all nonzero values improve over default, so no fitted parameter is renamed as a prediction. Score 2 reflects one minor non-load-bearing self-citation; no circular step.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The only free parameter is alpha. The main domain assumptions are about blank audio being a faithful no-audio condition and about the logit contrast isolating grounding; both are reasonable but untested. No new entities are postulated.

free parameters (1)
  • alpha = 1.0 (reported sweet spot; 0.5 also improves)
    Contrastive strength hyperparameter in Eq. 2. Selected by inspecting test-set F1 on the hallucination benchmark and Clotho-AQA (Section V-A, Table II); no held-out validation split.
assumptions (3)
  • domain assumption The difference between real-audio and blank-audio logits isolates audio-grounded evidence.
    Central to Eq. 2; assumes contrastive decoding amplifies tokens that depend on audio content and that zero-filled audio is a neutral baseline. Stated in Section II-B but not proven or stress-tested.
  • domain assumption LALM object hallucination is primarily driven by the model's prior yes bias, which blank-audio logits capture.
    Motivates the method in Section I and is supported indirectly by the ablation in Table II, where increasing alpha reduces the yes ratio from 82.6% to 39.0%.
  • standard math Softmax of a linear combination of logits is a valid decoding distribution.
    Standard contrastive decoding form from prior work; used in Eq. 2 without derivation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reducing Object Hallucination in Large Audio-Language Models via Audio-Aware Decoding." pith.science (2026). https://pith.science/paper/ZRIENHZE

@misc{pith2026250607233,
  author       = {Pith},
  title        = {Pith review of: Reducing Object Hallucination in Large Audio-Language Models via Audio-Aware Decoding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZRIENHZE}},
  note         = {Machine review of arXiv:2506.07233}
}
read the original 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.

Figures

Figures reproduced from arXiv: 2506.07233 by the authors.

Figure 1
Figure 1. Illustration of our proposed Audio-Aware Decoding (AAD). Since the [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Adaptive Perturbation Selection for Contrastive Audio Decoding

    cs.SD 2026-06 unverdicted novelty 6.0 of 10

    A learned per-example router over a 105-perturbation audio library improves contrastive decoding for audio-LLM hallucination, with task-dependent best distortions (e.g., reverse audio for temporal order).

Reference graph

Works this paper leans on

44 extracted references · 22 canonical work pages · cited by 1 Pith paper

  1. [23]

    Understanding sounds, missing the questions: The challenge of object hallucination in large audio-language models,

    C.-Y . Kuan, W.-P. Huang, and H.-y. Lee, “Understanding sounds, missing the questions: The challenge of object hallucination in large audio-language models,” in2024 Conference of the International Speech Communication Association (INTERSPEECH), 2024

    work page 2024

  2. [1]

    Gpt-4 technical report,

    OpenAI, “Gpt-4 technical report,” 2024. [Online]. Available: https: //arxiv.org/abs/2303.08774

    arXiv 2024

  3. [2]

    Instruction tuning with gpt-4,

    B. Peng, C. Li, P. He, M. Galley, and J. Gao, “Instruction tuning with gpt-4,” 2023. [Online]. Available: https://arxiv.org/abs/2304.03277

    arXiv 2023

  4. [3]

    Qwen2 technical report,

    A. Yanget al., “Qwen2 technical report,” 2024. [Online]. Available: https://arxiv.org/abs/2407.10671

    arXiv 2024

  5. [4]

    The llama 3 herd of models,

    A. Grattafioriet al., “The llama 3 herd of models,” 2024. [Online]. Available: https://arxiv.org/abs/2407.21783

    arXiv 2024

  6. [5]

    Gemini: A family of highly capable multimodal models,

    G. Teamet al., “Gemini: A family of highly capable multimodal models,” 2025. [Online]. Available: https://arxiv.org/abs/2312.11805

    arXiv 2025

  7. [6]

    A survey on multimodal large language models,

    S. Yin, C. Fu, S. Zhao, K. Li, X. Sun, T. Xu, and E. Chen, “A survey on multimodal large language models,”National Science Review, vol. 11, no. 12, Nov. 2024. [Online]. Available: http: //dx.doi.org/10.1093/nsr/nwae403

    doi:10.1093/nsr/nwae403 2024

  8. [7]

    The revolution of multimodal large language models: A survey,

    D. Caffagni, F. Cocchi, L. Barsellotti, N. Moratelli, S. Sarto, L. Baraldi, L. Baraldi, M. Cornia, and R. Cucchiara, “The revolution of multimodal large language models: A survey,” inFindings of the Association for Computational Linguistics: ACL 2024, L.-W. Ku, A. Martins, and V . Srikumar, Eds. Bangkok, Thailand: Association for Computational Linguistics...

    work page 2024

Show all 44 references
  1. [8]

    MM-LLMs: Recent advances in MultiModal large language models,

    D. Zhang, Y . Yu, J. Dong, C. Li, D. Su, C. Chu, and D. Yu, “MM-LLMs: Recent advances in MultiModal large language models,” inFindings of the Association for Computational Linguistics: ACL 2024, L.-W. Ku, A. Martins, and V . Srikumar, Eds. Bangkok, Thailand: Association for Co...

    2024

  2. [9]

    Audiopalm: A large language model that can speak and listen,

    P. K. Rubensteinet al., “Audiopalm: A large language model that can speak and listen,” 2023. [Online]. Available: https://arxiv.org/abs/2306. 12925

    2023

  3. [10]

    Joint audio and speech understanding,

    Y . Gong, A. H. Liu, H. Luo, L. Karlinsky, and J. Glass, “Joint audio and speech understanding,” in2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), 2023, pp. 1–8

    2023

  4. [11]

    Desta: Enhancing speech language models through descriptive speech-text alignment,

    K.-H. Lu, Z. Chen, S.-W. Fu, H. Huang, B. Ginsburg, Y .-C. F. Wang, and H. yi Lee, “Desta: Enhancing speech language models through descriptive speech-text alignment,” inInterspeech 2024, 2024, pp. 4159– 4163

    2024

  5. [12]

    Developing instruction-following speech language model without speech instruction-tuning data,

    K.-H. Lu, Z. Chen, S.-W. Fu, C.-H. H. Yang, J. Balam, B. Ginsburg, Y .-C. F. Wang, and H.-Y . Lee, “Developing instruction-following speech language model without speech instruction-tuning data,” inICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signa...

    2025

  6. [14]

    Audio flamingo 2: An audio-language model with long-audio understanding and expert reasoning abilities,

    S. Ghosh, Z. Kong, S. Kumar, S. Sakshi, J. Kim, W. Ping, R. Valle, D. Manocha, and B. Catanzaro, “Audio flamingo 2: An audio-language model with long-audio understanding and expert reasoning abilities,”

  7. [15]

    On the landscape of spoken language models: A comprehensive survey,

    S. Arora, K.-W. Chang, C.-M. Chien, Y . Peng, H. Wu, Y . Adi, E. Dupoux, H.-Y . Lee, K. Livescu, and S. Watanabe, “On the landscape of spoken language models: A comprehensive survey,”arXiv preprint arXiv:2504.08528, 2025

    2025 arXiv

  8. [16]

    Dynamic-superb: Towards a dynamic, collaborative, and comprehensive instruction-tuning benchmark for speech,

    C.-Y . Huanget al., “Dynamic-superb: Towards a dynamic, collaborative, and comprehensive instruction-tuning benchmark for speech,” inICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024

    2024

  9. [17]

    Dynamic-superb phase-2: A collabo- ratively expanding benchmark for measuring the capabilities of spoken language models with 180 tasks,

    C.-Y . Huang, W.-C. Chenet al., “Dynamic-superb phase-2: A collabo- ratively expanding benchmark for measuring the capabilities of spoken language models with 180 tasks,” inICLR, 2025

    2025

  10. [18]

    MMAU: A massive multi- task audio understanding and reasoning benchmark,

    S. Sakshi, U. Tyagi, S. Kumar, A. Seth, R. Selvakumar, O. Nieto, R. Duraiswami, S. Ghosh, and D. Manocha, “MMAU: A massive multi- task audio understanding and reasoning benchmark,” inThe Thirteenth International Conference on Learning Representations, 2025

    2025

  11. [19]

    Sakura: On the multi-hop reasoning of large audio-language models based on speech and audio information,

    C.-K. Yang, N. Ho, Y .-T. Piao, and H. yi Lee, “Sakura: On the multi-hop reasoning of large audio-language models based on speech and audio information,” 2025. [Online]. Available: https://arxiv.org/abs/2505.13237

    2025 arXiv

  12. [20]

    Audiobench: A universal benchmark for audio large language models,

    B. Wang, X. Zou, G. Lin, S. Sun, Z. Liu, W. Zhang, Z. Liu, A. Aw, and N. F. Chen, “Audiobench: A universal benchmark for audio large language models,”NAACL, 2025

    2025

  13. [21]

    V oicebench: Benchmarking llm-based voice assistants,

    Y . Chen, X. Yue, C. Zhang, X. Gao, R. T. Tan, and H. Li, “V oicebench: Benchmarking llm-based voice assistants,”arXiv preprint arXiv:2410.17196, 2024

    2024 arXiv

  14. [22]

    Towards holistic evaluation of large audio-language models: A comprehensive survey,

    C.-K. Yang, N. S. Ho, and H.-y. Lee, “Towards holistic evaluation of large audio-language models: A comprehensive survey,”arXiv preprint arXiv:2505.15957, 2025

    2025 arXiv

  15. [24]

    Can large audio-language models truly hear? tackling hallucinations with multi-task assessment and stepwise audio reasoning,

    C.-Y . Kuan and H.-Y . Lee, “Can large audio-language models truly hear? tackling hallucinations with multi-task assessment and stepwise audio reasoning,” inICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025, pp. 1–5

    2025

  16. [25]

    A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions,

    L. Huang, W. Yu, W. Ma, W. Zhong, Z. Feng, H. Wang, Q. Chen, W. Peng, X. Feng, B. Qinet al., “A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions,” ACM Transactions on Information Systems, vol. 43, no. 2, pp. 1–55, 2025

    2025

  17. [26]

    Siren’s song in the ai ocean: A survey on hallucination in large language models,

    Y . Zhang, Y . Li, L. Cui, D. Cai, L. Liu, T. Fu, X. Huang, E. Zhao, Y . Zhang, Y . Chen, L. Wang, A. T. Luu, W. Bi, F. Shi, and S. Shi, “Siren’s song in the ai ocean: A survey on hallucination in large language models,” 2023. [Online]. Available: https://arxiv.org/abs/2309.01219

    2023 arXiv

  18. [27]

    A survey of hallucination in large foundation models,

    V . Rawte, A. Sheth, and A. Das, “A survey of hallucination in large foundation models,” 2023. [Online]. Available: https: //arxiv.org/abs/2309.05922

    2023 arXiv

  19. [28]

    A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions,

    L. Huanget al., “A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions,”ACM Transactions on Information Systems, vol. 43, no. 2, p. 1–55, Jan

  20. [29]

    Speech-ifeval: Evaluating instruction-following and quantifying catastrophic forgetting in speech- aware language models,

    K.-H. Lu, C.-Y . Kuan, and H.-y. Lee, “Speech-ifeval: Evaluating instruction-following and quantifying catastrophic forgetting in speech- aware language models,”arXiv preprint arXiv:2505.19037, 2025

    2025 arXiv

  21. [30]

    Available: http://dx.doi.org/10.1145/3703155

    [Online]. Available: http://dx.doi.org/10.1145/3703155

    doi

  22. [31]

    Retrieval- augmented generation for knowledge-intensive nlp tasks,

    P. Lewis, E. Perez, A. Piktus, F. Petroni, V . Karpukhin, N. Goyal, H. K ¨uttler, M. Lewis, W.-t. Yih, T. Rockt ¨aschelet al., “Retrieval- augmented generation for knowledge-intensive nlp tasks,”Advances in neural information processing systems, vol. 33, pp. 9459–9474, 2020

    2020

  23. [32]

    Trusting your evidence: Hallucinate less with context-aware decoding,

    W. Shi, X. Han, M. Lewis, Y . Tsvetkov, L. Zettlemoyer, and W.-t. Yih, “Trusting your evidence: Hallucinate less with context-aware decoding,” inProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Te...

    2024

  24. [33]

    Contrastive decoding: Open-ended text generation as optimization,

    X. L. Li, A. Holtzman, D. Fried, P. Liang, J. Eisner, T. Hashimoto, L. Zettlemoyer, and M. Lewis, “Contrastive decoding: Open-ended text generation as optimization,” inProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Paper...

    2023

  25. [34]

    Retrieval-augmented generation for large language models: A survey,

    Y . Gao, Y . Xiong, X. Gao, K. Jia, J. Pan, Y . Bi, Y . Dai, J. Sun, Q. Guo, M. Wanget al., “Retrieval-augmented generation for large language models: A survey,”CoRR, 2023

    2023

  26. [35]

    Qwen2-audio technical report,

    Y . Chu, J. Xu, Q. Yang, H. Wei, X. Wei, Z. Guo, Y . Leng, Y . Lv, J. He, J. Lin, C. Zhou, and J. Zhou, “Qwen2-audio technical report,”

  27. [36]

    Salmonn: Towards generic hearing abilities for large language models,

    C. Tang, W. Yu, G. Sun, X. Chen, T. Tan, W. Li, L. Lu, M. Zejun, and C. Zhang, “Salmonn: Towards generic hearing abilities for large language models,” inThe Twelfth International Conference on Learning Representations

  28. [37]

    Delta – contrastive decoding mitigates text hallucinations in large language models,

    C. P. Huang and H.-Y . Chen, “Delta – contrastive decoding mitigates text hallucinations in large language models,” 2025. [Online]. Available: https://arxiv.org/abs/2502.05825

    2025 arXiv

  29. [38]

    Contrastive decoding: Open- ended text generation as optimization,

    X. L. Li, A. Holtzman, D. Fried, P. Liang, J. Eisner, T. Hashimoto, L. Zettlemoyer, and M. Lewis, “Contrastive decoding: Open- ended text generation as optimization,” 2023. [Online]. Available: https://arxiv.org/abs/2210.15097

    2023 arXiv

  30. [39]

    Clotho-aqa: A crowdsourced dataset for audio question answering,

    S. Lipping, P. Sudarsanam, K. Drossos, and T. Virtanen, “Clotho-aqa: A crowdsourced dataset for audio question answering,” in2022 30th European Signal Processing Conference (EUSIPCO), 2022, pp. 1140– 1144

    2022

  31. [40]

    AudioCaps: Generating captions for audios in the wild

    C. D. Kim, B. Kim, H. Lee, and G. Kim, “AudioCaps: Generating captions for audios in the wild.” Association for Computational Linguistics, 2019, pp. 119–132. [Online]. Available: https://aclanthology. org/N19-1011/

    2019

  32. [41]

    Clotho-aqa: A crowdsourced dataset for audio question answering,

    S. Lipping, P. Sudarsanam, K. Drossos, and T. Virtanen, “Clotho-aqa: A crowdsourced dataset for audio question answering,” 2022. [Online]. Available: https://arxiv.org/abs/2204.09634

    2022 arXiv

  33. [42]

    Finetuned language models are zero-shot learners,

    J. Wei, M. Bosma, V . Zhao, K. Guu, A. W. Yu, B. Lester, N. Du, A. M. Dai, and Q. V . Le, “Finetuned language models are zero-shot learners,” inInternational Conference on Learning Representations, 2022

    2022

  34. [45]

    On the audio hallucinations in large audio-video language models,

    T. Nishimura, S. Nakada, and M. Kondo, “On the audio hallucinations in large audio-video language models,” 2024. [Online]. Available: https://arxiv.org/abs/2401.09774

    2024 arXiv

  35. [2024]

    Available: https://arxiv.org/abs/2407.10759

    [Online]. Available: https://arxiv.org/abs/2407.10759

    arXiv

  36. [2025]

    Available: https://arxiv.org/abs/2503.03983

    [Online]. Available: https://arxiv.org/abs/2503.03983

    arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.