A new audit says the fix for AI that hears sounds that aren't there works, but not for the reason its creators assumed.
Researchers built a black-box causal test to check whether audio-aware decoding (AAD), the standard technique for curbing hallucinations in large audio-language models, actually uses the audio evidence it claims to. AAD contrasts model outputs against a blank, all-zero waveform; prior critics suspected that comparison works only because a blank clip is wildly out-of-distribution, not because it isolates real evidence. The team instead localized the specific evidence for a claim using CLAP similarity, removed only that region with spectral gating, and compared the result against randomly removing a same-length chunk, across three models. On Qwen2-Audio-7B, that targeted removal alone scored lower (AUROC 0.613) than the blank-waveform swap (0.721), but once clip duration was held constant, localized removal beat random removal in every test, showing the gap was a duration artifact rather than proof that AAD ignores the real evidence.
Combining that localized evidence with decoding is what actually moves the needle. The researchers built a contrastive decoding method that gates and ensembles both blank-waveform and language-prior contrasts together with the localized evidence, lifting Qwen2-Audio's accuracy from 57.1% to 71.9% and beating plain AAD. Yet the causal grounding pattern didn't carry over to Audio Flamingo 3, so there's no single grounding recipe that explains hallucination across models.
In other words, audio hallucination isn't one bug with one fix. It's a different bug on every model, and anyone pitching a universal patch is getting ahead of the data.