All 10 LALMs hallucinate on adversarial Absent/Contradicted Premise audio questions (36–74% rates)

Auto-published from arXiv:2609.33893 by the mine-arxiv pipeline. Reviewed by an LLM judge (Sonnet) against the archive bar — see CONTRIBUTING. Notes: cleared review (confidence 0.78, flags: [no-prompt-excerpt])

Category

hallucination

Model

Qwen2-Audio-Instruct, Qwen2.5-Omni, Kimi-Audio-Instruct, Audio-Flamingo-3, Step-Audio-2-mini, MiMo-Audio-Instruct, Qwen3-Omni-Instruct, Covo-Audio, MiMo-V2.5, Gemini 3.7 Flash

Surface

LALM API / audio-language model input (audio clip + text question)

Setup

MISHAP-Bench constructs adversarial open-ended questions using seven strategies. For context hallucination: 'Contradicted Premise' asks about an audio event that contradicts actual content (e.g., 'how many drum-kit hits are audible' when the audio contains only fingerstyle guitar); 'Absent Premise' asks about entities not in the clip at all (e.g., questions about 'cat meows in a doorbell-only clip'). No verbatim prompt template quoted in main text; examples appear in Figure 7 (appendix). 12,000 question-audio pairs across 3,000 clips spanning music, sound, and speech domains.

Reproduction anchor

code: https://anonymous.4open.science/w/MISHAP-Bench/ ; benchmark: MISHAP-Bench (3,000 audio clips from Song Describer Dataset, Clotho v2.1, FSD50K, LibriSpeech, Common Voice 22.0; full question set and evaluation pipeline in repo); judge prompts in Figures 8–9 (appendix)

Observed behavior

All ten models produce ungrounded claims affirming or elaborating on the false/absent premise rather than identifying the premise as invalid. Hallucination rates range from 36.1% (Qwen2.5-Omni) to 74.8% (MiMo-Audio-Instruct). Even the best-performing model (Gemini 3.7 Flash, 36.5%) hallucinated on more than one-third of adversarial context questions. No model approached a low hallucination rate.

Expected behavior

A correctly-behaving LALM should detect that the question's premise contradicts or is absent from the audio, state that the claimed event does not appear in the clip, and refuse to elaborate on the false premise.

Reproducibility

high

Threat model

Developers deploying LALMs in audio analysis, accessibility, or content moderation pipelines may receive fluent but factually ungrounded answers whenever a user (maliciously or accidentally) frames a question with a false premise about the audio. Automated pipelines that do not verify model outputs would propagate hallucinated descriptions as ground truth.

Novelty

First benchmark to systematically distinguish 'context hallucination' (audio-ungrounded claims) from 'knowledge hallucination' in audio-language models using adversarial question-generation strategies, revealing that all ten state-of-the-art LALMs fail at rates that make production deployment unreliable.

Source

• arXiv: 2609.33893
• PDF: https://arxiv.org/pdf/2609.33893
• Categories: cs.LG
• Authors: Zhi Wen Soi, Giulio Segalini, Jian-Jia Chen, Lydia Chen

Triage notes (auto)

• paperType: benchmark
• estimatedCaseCount: 3
• triage reason: Comprehensive hallucination benchmark with 12,000 test cases demonstrating context and knowledge hallucinations across 10 LALMs; documents concrete failures (e.g., Gemini 3.7 Flash 36.5% hallucination rate).

─────────────────────────────────────────────────────────────────────