acceptodds
Under review as a conference paper at ICLR 2027

MISHAP-Bench: A Hallucination Benchmark for Large Audio-Language Models

Abstract

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.

Primary area:
datasets and benchmarks
Keywords:
large audio-language models, hallucination benchmark

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ueQnOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents MISHAP-Bench, a benchmark designed to measure hallucinations in large audio-language models (LALMs). Its evaluation focuses on whether generated claims are supported by the input audio or by externally verifiable evidence, rather than on answer correctness alone. The benchmark includes 3,000 audio clips from music, sound, and speech, as well as 12,000 open-ended questions covering two categories: context hallucination and knowledge hallucination. To assess model responses, the authors develop a groundedness judge that combines question-specific reference rubrics, human annotations, category-specific prompts, and web-based evidence retrieval. They evaluate ten LALMs and four inference-time hallucination mitigation methods. The results indicate that current models still hallucinate frequently: even Gemini 3.7 Flash exhibits a hallucination rate of 36.5%. Among the mitigation approaches, AAD achieves the most consistent improvements, although its effectiveness varies across models, audio domains, and hallucination types. Overall, MISHAP-Bench provides a relevant and potentially useful resource for studying the reliability of LALMs. However, the benchmark contains an uneven distribution of audio clip durations, which may affect the comparability of results across samples and the reliability of the evaluation. The role and processing procedure of the 1,000 human-annotated responses are not described in sufficient detail. The paper explains that these annotations are used to refine the reference rubrics and judge prompts, select in-context examples, and assess judge-human agreement, but it does not clearly specify how the annotated data are separated between prompt development, example selection, and independent evaluation. As a result, the reported agreement between the groundedness judge and human annotators may be affected by a potential calibration-evaluation overlap, and the extent to which the judge generalizes to the remaining samples is difficult to assess.

Critical strengths
  1. The paper's main contribution is decoupling answer correctness from whether a response's claims are actually supported by the audio or external evidence. This is well-motivated for open-ended audio QA, where a partially-correct answer may carry unsupported details, and where refusing a false premise or expressing uncertainty is the reliable behavior. The shift closely matches what users actually care about in LALM deployment, and is especially apt given the heterogeneity of audio evidence.

  2. Separating audio-grounded hallucination from audio-related external-knowledge hallucination explains where failures originate. The authors show the two metrics do not always move together, evidencing that a single hallucination rate masks distinct behavioral modes.

  3. Per-clip generated questions with three context strategies and four knowledge strategies probe premise acceptance, assertion of unknowables, and appropriate abstention. Treating premise-correction and uncertainty-expression as grounded is a valuable correction to correctness-based scoring.

  4. The paper not only includes a benchmark but also compares existing hallucination mitigation methods, thus expanding its academic value. The results expose trade-offs: higher context hallucination on music and sound, better context grounding but persistent knowledge hallucination on speech, and AAD generally more robust than VCD/MTI/DoLa yet still degrading in places.

Critical weaknesses
  1. The role and processing procedure of the 1,000 human-annotated responses are not described in sufficient detail. The paper explains that these annotations are used to refine the reference rubrics and judge prompts, select in-context examples, and assess judge–human agreement, but it does not clearly specify how the annotated data are separated between prompt development, example selection, and independent evaluation. As a result, the reported agreement between the groundedness judge and human annotators may be inflated by a potential calibration–evaluation overlap, and the extent to which the judge generalizes to the remaining samples is difficult to assess.

  2. The clips come from six existing datasets and inherit their domain, recording-condition, and language biases. Because four questions share the same audio and description, the effective independent sample size is closer to 3,000 than 12,000. The dataset's actual representativeness is therefore lower than claimed in the paper.

  3. The duration distribution in the music domain is unbalanced and inconsistent with the other two domains. Unlike speech and sound, whose clip lengths are comparatively reasonable, the music clips concentrate on extremely short segments (under 1 s) and ultra-long segments (around 2 min), omitting the 10–30 s range. This skew makes cross-domain comparison unreliable, since a model's music hallucination rate may be driven by degenerate very-short or very-long inputs rather than by genuine difficulty.

  4. The hallucination-mitigation experiments read more like an supplementary material than a native contribution of the benchmark.

Questions for authors
  1. Why is the duration distribution in the music domain so unbalanced compared to speech and sound?

  2. Could you explain in detail the role that the hallucination-mitigation experiments play in supporting the paper's main contribution?

Reviewer_1DkqOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces MISHAP-Bench, comprising 12,000 open-ended questions over 3,000 audio clips spanning music, sound, and speech. It distinguishes context hallucinations, involving claims unsupported by the audio, from knowledge hallucinations, involving unsupported external facts. Responses are evaluated using an LLM judge with question-specific rubrics, human-guided prompts, and web search. The authors evaluate ten LALMs and compare four inference-time mitigation methods, finding substantial hallucination and uneven mitigation gains.

Critical strengths
  1. Relevant motivation. Evaluating groundedness beyond answer correctness addresses an important limitation: a response can contain unsupported details even when its main answer is correct.

  2. Useful diagnostic coverage. The context/knowledge distinction, seven question strategies, and domain-level analyses provide a structured examination of model failures.

  3. Broad and well-documented evaluation. The paper evaluates ten models and compares four mitigation methods under a common protocol. Prompts, model settings, and mitigation hyperparameters are documented, and judge-component ablations provide useful supporting analysis.

Critical weaknesses
  1. Independent validation of the judge is unclear. Human annotations are used to refine rubrics, revise prompts, select examples, and choose the final formulation, but a separate held-out set for reporting agreement is not documented. This may overestimate reliability. Validation should also cover Gemini 3.7 Flash and MiMo-V2.5. The number of annotators per item, disagreement-resolution procedure, and sample underlying human–human agreement need clarification.

  2. Some judge rules conflict with the stated groundedness definition. Figure 9 allows a response to challenge a false premise while supplying an incorrect replacement fact and still be labeled grounded. Figure 14 similarly accepts a “plausible or correct year.” These rules appear to permit unsupported factual claims. The authors should justify these exceptions and clarify whether all returned reasoning is evaluated alongside the final answer.

  3. Description-based judging may misclassify audio-grounded claims. The context judge treats the description as a complete record and assumes unmentioned content is absent. However, captions, transcripts, and metadata do not exhaustively capture audible events, speaker characteristics, or musical attributes. Although acknowledged as a limitation, this directly affects evaluation validity. An audio-based human audit is needed to quantify the resulting errors.

  4. The source of knowledge hallucination is unclear. Responses to fabricated sources or false factual premises may reflect failures of the underlying LLM rather than audio processing. Question-only and transcript- or description-conditioned baselines would help establish how much the audio contributes to these errors and distinguish general factual hallucination from failures of acoustic grounding.

  5. Mitigation gains are not sufficiently separated from reduced usefulness. The paper reports abstention on knowledge questions, but does not measure unnecessary refusal on answerable questions or show how mitigation changes abstention. Answerable controls and general audio-understanding accuracy before and after mitigation, for example on MMAU or MMAR, would clarify whether improvements preserve useful capabilities. Uncertainty estimates are also needed to interpret small differences such as 36.1 versus 35.9.

  6. Related work needs correction and clearer positioning. Table 7 characterizes HalluAudio (Zhao et al., ACL 2026) as binary/MCQ, although it also includes open-ended QA, mixed-audio conditions, and refusal-rate analysis. Kuan & Lee’s Can Large Audio-Language Models Truly Hear? is cited, but the distinction from its evaluation and mitigation setting could be clearer. Other directly relevant work is missing:

    These comparisons would clarify MISHAP’s contribution and the scope of its conclusions, since its benchmark questions introduce misleading information through text rather than audio-side attacks.

Questions for authors
  1. Can you report judge–human agreement on a held-out annotated set covering all ten models? How were annotation disagreements resolved, and were calibration examples excluded from judge-validation results?
  2. Why do some judge rules accept incorrect replacement facts or merely plausible answers as grounded? Does evaluation include all returned reasoning?
  3. How often does incomplete audio-description evidence cause incorrect verdicts, according to an audit with access to the audio?
  4. How do knowledge-hallucination rates compare across audio-conditioned, question-only, and transcript- or description-conditioned settings?
  5. Do mitigation improvements persist when accounting for false refusals, general audio-understanding accuracy, and statistical uncertainty?
Reviewer_SBDROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses hallucination evaluation for large audio-language models (LALMs). The authors argue that existing audio hallucination benchmarks, which use closed-form questions scored by answer correctness, have two shortcomings: a response may be marked correct despite containing unsupported details, and a response that expresses uncertainty or corrects a flawed premise may be marked wrong despite being well supported. Correctness is therefore not a reliable indicator of fabrication. The authors instead define hallucination in terms of whether a claim is backed by evidence, distinguishing context hallucination (claims unsupported by the audio) from knowledge hallucination (claims about audio-related topics unsupported by external facts).

On this basis, they construct MISHAP-Bench, comprising 3,000 clips (1,000 each for music, sound, and speech, drawn from two public datasets per domain) and 12,000 open-ended questions. Questions are generated by Nemotron 3 Nano Omni following seven hallucination-inducing strategies (e.g., Absent Premise, False Authority, Unknowable Provenance), each paired with a reference rubric produced by MiniMax M3 that specifies which responses count as hallucinated and which as grounded. Judgments are made by a text-only judge based on Qwen3.8-27B, which receives the question, model response, audio description, and rubric, with web search enabled for knowledge questions. The rubrics and judge prompts are refined through repeated comparison with 1,000 human-annotated responses, reaching κ=0.80 and 90.1% agreement on that subset.

Experiments on ten LALMs show that all models hallucinate at substantial rates (the best at roughly 36%), with music being the most challenging domain and speech the least; a stance analysis indicates that the lower knowledge-hallucination rates of some models coincide with higher abstention. The authors further unify four contrastive decoding methods (AAD, VCD, MTI, DoLa) under a single formulation and evaluate them on the eight open-weight models, finding that only the audio-specific AAD improves all models, while the others trade gains in some domains for losses in others.

Critical strengths
  1. Shifting audio hallucination evaluation from answer correctness to response-level evidential support is a substantive improvement over the existing paradigm. Existing audio hallucination benchmarks rely on closed-form questions and label matching, so their unit of evaluation is the answer rather than the response. This paper centers evaluation on groundedness: each claim in an open-ended response is checked against evidence, and fabrication is explicitly distinguished from several forms of grounded behavior (direct answers, abstention, premise correction, explicitly labeled guesses). Context questions are verified against the audio description, while knowledge questions can additionally be checked against external facts via web search. As a result, the paper offers not merely a new dataset but a transferable evaluation protocol; even if the judge’s specific implementation can be improved, the framework itself remains a useful reference for future work.

  2. Evaluation and mitigation are conducted under a single protocol, with sufficient coverage and transparency to support reuse by the community. The paper evaluates ten LALMs and four contrastive decoding methods under the same groundedness-based evaluation, casting the latter into a unified formulation with fixed hyperparameters so that comparisons across methods and models are interpretable. The experiments surface several informative patterns, such as the consistent failure of all models on context questions in the music and sound domains, and the observation that similar overall HR can correspond to markedly different response behaviors. These findings go beyond a leaderboard and point to concrete failure modes and directions for improvement. The appendix provides the full judge prompts, rubric templates, model checkpoints, decoding configurations, mitigation hyperparameters, and complete result tables, along with the human annotation interface, offering ample detail for reproduction and extension.

Critical weaknesses
  1. The groundedness judge’s reliance on audio-description completeness may affect evaluation reliability.

The groundedness judge operates on source-provided audio descriptions rather than the raw audio. Appendix A.3 (Figure 8) explicitly instructs the judge to treat any content absent from the description as absent from the clip. Since these descriptions may omit genuine background sounds or acoustic details, content that a model correctly perceives could be misclassified as hallucination. Because the same descriptions are also used for question generation and reference-rubric construction, such omissions may further propagate into question premises and evaluation criteria.

The reported judge–human agreement (90.1%) is encouraging, but human annotators were also shown the same audio description and reference rubric. The current agreement results therefore may not fully rule out a systematic bias introduced by shared reference material. An independent human evaluation based on the raw audio alone, without access to the description or rubric, would help quantify the actual impact of description omissions on hallucination judgments.

  1. Hallucination Rate does not fully capture the distinction between justified abstention and informative answers.

The paper adopts Hallucination Rate (HR) as its primary metric, under which both justified abstentions and correct, well-grounded answers are counted as grounded. A model could therefore lower its HR by abstaining more often, without necessarily improving its ability to provide correct and useful answers. While abstention is appropriate when the available evidence is insufficient, over-abstention on answerable questions may yield favorable HR at the cost of informativeness.

The authors analyze abstention in Figure 6 and acknowledge that lower knowledge-hallucination rates do not imply stronger factual knowledge. Nevertheless, the overall HR may not adequately reflect the trade-off between reliability and answer utility. In addition, whether the HR reductions achieved by mitigation methods such as AAD stem from more accurate factual judgment or from more frequent abstention has not been fully verified.

  1. The dependence of knowledge questions on audio input has not been sufficiently validated.

Knowledge questions constitute 50% of the benchmark, yet several examples in Figures 7 and 14 (e.g., those involving historical dates, false geographic premises, or fabricated sources) appear answerable from the question text and external knowledge alone, without listening to the audio. The hallucination rate for this category may therefore largely reflect the underlying LLM’s factual knowledge and false-premise detection rather than its use of audio information.

Although the authors explicitly scope audio-related knowledge hallucination into the benchmark, knowledge and context questions are weighted equally in the overall HR, which may affect how the aggregate metric should be interpreted with respect to audio-related reliability. Question-only and description-only baselines would help quantify the actual contribution of audio input to knowledge questions and clarify the rationale for combining the two categories into a single score.

  1. The independence between judge development and evaluation data is unclear.

In §3.3, 1,000 human-annotated responses are used to iteratively refine the reference rubrics and judge prompts, and in-context examples are drawn from the same set. The final judge–human agreement (90.1%, κ≈0.80) is also reported on this annotated subset, and the paper does not state whether a separate development/test split was used. The reported agreement may therefore be affected by prompt-optimization bias.

Figure 5 shows hallucination rates close to human estimates on two external benchmarks, but this evidence has some limitations. First, AHa-Bench and HalluAudio are primarily closed-form tasks with human hallucination rates of 15.5% and 9.6%, which differ substantially in task format and hallucination-rate distribution from MISHAP-Bench’s open-ended adversarial setting (57.4%). The latter requires the judge to handle multiple factual claims, hedged statements, and false premises, and the external evaluation does not appear to validate the judge’s reliability in these scenarios. Second, neither external benchmark covers knowledge hallucination, so the knowledge judge, which was designed and tuned for MISHAP-Bench-specific strategies, is not validated by this comparison. Third, Figure 5 reports only aggregate hallucination rates without per-sample agreement metrics.

Taken together, the current validation may not fully exclude the risk of the judge overfitting to the development samples.

Questions for authors

Q1 Are the samples used to refine the rubric templates and judge prompts the same 1,000 annotated responses used to report κ=0.80 / 90.1% agreement? If a development/test split was used, please state the split ratio and report agreement on the test portion alone. If no split was used, could the authors draw an independent sample from MISHAP-Bench after the rubrics and prompts were frozen and report per-sample κ and agreement on it? Please also clarify how many annotators labeled each sample, how disagreements were resolved, and how κ_hh=0.88 was computed.

Q2 Among the music/sound context responses that the judge labeled as hallucinated, was any subset re-checked by annotators using only the raw audio, without access to the description or rubric? If so, what fraction turned out to be misclassified? If not, could the authors add such a check on a small subset (e.g., 200 responses)?

Q3 Could the authors report, for each model, the proportion of knowledge responses that are both asserted and grounded, together with the effect of the four mitigation methods on abstention rates? In particular, AAD reduces Kimi-Audio’s HR on False Authority from 56.0 to 28.3; how much of this reduction corresponds to an increase in abstention?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.