Erase the audio region an LALM points to and its "yes" often flips: 16.8% versus 3.1% for a random window

A hallucinated sound claim in an audio-language model can be tested by deleting the audio region the model itself points to: if the claim survives that deletion, it was not grounded in that evidence. The prior inference-time mitigation, Audio-Aware Decoding, contrasts the real recording against an all-zero blank waveform, which is out-of-distribution for the audio encoder and conflates absent evidence with malformed input. On Qwen2-Audio-7B, ablating the CLAP-localized window gives raw AUROC 0.613 against 0.721 for blanking; but blanking silences the whole clip (median about 10 s) while localized windows span 2 to 4 s, and a duration-only regression fit on random-window cells predicts blank's mean effect almost exactly (0.126 predicted versus 0.126 actual), so the raw edge is largely a duration artifact.

Once the amount removed is held fixed, localized removal beats an equally sized random window at every grid cell on both models. Earlier comparisons mixed different intervention sizes, so whether localization helps could not be read off the raw AUROC ranking. At 4 s and 30 dB, mean drop is 0.079 for localized versus 0.016 for random on Qwen2-Audio (Wilcoxon significant) and 0.126 versus 0.016 on Audio Flamingo 3; in the synthetic control, ablating the true insertion window flips 16.8% of true positives, CLAP's pick 10.5%, and a random window 3.1%, with mean CLAP-to-true-location IoU of 0.45.

The same contrast also corrects answers at decoding time: a language-prior contrast lifts Qwen2-Audio accuracy from 57.1% to 68.2%, and combining it with the blank-audio contrast under gating reaches 71.9%. Blank-audio contrastive decoding reaches only 64.7% on the same model, and the localized contrast alone reaches 59.6%. On Qwen2-Audio the language-prior contrast fixes some hallucinations while breaking some correct claims, and AAD-CD is more conservative, with only partially overlapping errors, so the combination exceeds either single contrast's own best; 5-fold cross-validation gives about 14.8 percentage points.

The test targets object-existence judgments in audio-language models, runs on the object-existence split of the public AudioCaps, ESC-50 and VocalSound benchmark, and covers three models, Qwen2-Audio-7B, Audio Flamingo 3 and Qwen2.5-Omni-7B, all queried black-box with no training. The combined and gated decoding rule is validated on Qwen2-Audio, needs two forward passes per query, and treats the language prior as a precomputed per-class constant.

On Qwen2-Audio the localized-versus-random gap is only partly a CLAP-accuracy artifact, and even ablating the true location flips only 16.8% of these claims, so genuine partial location-insensitivity survives. The language-prior ensemble is significant per example but not per class, and no multiple-comparison correction is applied. Qwen2.5-Omni has only headline numbers, and Phi-4-multimodal and FLAM remain non-functional due to dependency incompatibilities. On the shorter HalluAudio benchmark the matched-duration effect attenuates sharply, though a small residual survives on Qwen2-Audio. Cross-model transfer of the causal profile loses in both directions to the target model's own best fixed contrast, which the authors treat as a genuine negative result.