Does Location Matter? A Black-Box Causal Audit Reveals Model-Dependent Audio Grounding in LALM Hallucination

• AAD contrasts logits from real audio with an all-zero “blank” waveform, which may conflate absent evidence with malformed input.Source: §1 · p. 1
• Localized evidence removal provides a black-box causal test, while sufficiency testing checks whether inserting a real exemplar flips a negative claim.The method localizes evidence with sliding-window CLAP similarity and removes it through spectral gating, leaving the rest of the real recording intact.Source: §1 · p. 1
• The test requires no gradients, internal activations, or training, unlike the cited concurrent causal approaches [8] [9] [10] [11].Source: §1 · p. 1
• Raw detection value remains model-dependent; on Qwen2-Audio-7B-Instruct, localized ablation’s AUROC trails blank’s because of a duration confound.Source: §1 · p. 1

• The object-existence benchmark [4] is used here, while its follow-up broadens evaluation to temporal-order and attribute hallucinations [5].Source: §2 · p. 1
• AAD [1] is the closest prior mitigation and serves as the baseline; other approaches calibrate against content-free or distorted inputs [14–17].Source: §2 · p. 1
• Prior causal-grounding approaches include training a reward [8], tracing layer effects [9], reading activations [10], or training a frame-level head [11].The paper’s tests instead intervene on raw audio without training.Source: §2 · p. 1

3.1. Necessity: localized spectral-gate ablation

• Sliding-window CLAP selects the highest-similarity evidence region, which is attenuated by spectral gating before the model is queried again.Windows are 2 or 4 seconds with a 0.5-second hop; gating uses −40 or −100 dB attenuation.Source: §3.1 · p. 1

3.2. Controls

• Random-window, blank-audio, language-only, and real-unrelated-audio conditions provide controls, while sufficiency inserts a real exemplar into a negative-control clip.Conditions are scored using deterministic, forced-choice “yes”/“no” probabilities.Source: §3.2 · pp. 1–2

3.3. Oracle-localization control

• The oracle-localization control compares ablation of the true insertion window, CLAP’s pick, and a random window on the same affirmed mixture.Temporal IoU between CLAP’s window and the known true location measures localization accuracy.Source: §3.3 · p. 2

3.4. From detection to mitigation: contrastive decoding

• Contrastive decoding reuses logits to compare blank-audio, language-prior, and localized-ablation contrasts while sweeping α from 0 to 3.The best rule uses two forward passes per query; the language prior is a precomputed per-class constant.Source: §3.4 · p. 2

3.5. Theoretical framing: from a binary PN to a graded curve

• The graded necessity curve records the minimum fraction of top-ranked windows whose removal flips a claim, with the estimate censored at 1.Paired localization efficiency compares CLAP-ranked removal with random-order removal; six checkpoints are used, and thresholded decisions are nearly monotone.Source: §3.5 · p. 2

3.6. A causal grounding profile, and its identifiability

• The causal-grounding profile combines five signals, while positive specificity rejects window exchangeability but does not establish identification.The interpretation of waveform deletion as do(Eo:=0) requires that the removed span be the only disturbed causal channel; CLAP salience bias could also yield positive specificity.Source: §3.6 · p. 2

• Three models are queried identically in a black-box setup, with Qwen2-Audio-7B-Instruct as the primary subject and Audio Flamingo 3 and Qwen2.5-Omni-7B as cross-model checks.Source: §4 · p. 2
• The object-existence split contains 10,800 rows across 53 classes; Qwen2-Audio’s 5,000-row sample includes 4,355 affirmative claims and a 43% overclaim rate.Audio Flamingo 3 and Qwen2.5-Omni use different sample sizes and affirmation rates.Source: §4 · p. 2
• For Audio Flamingo 3, 83 hallucinated cases yield per-cell AUROC values ranging from 0.37 to 0.58, so flip-rate is the reliable metric for this model.The reported metrics include AUROC/AUPRC and empirical PN/PS, with bootstrap 95% confidence intervals.Source: §4 · pp. 2–3

5.1. Blank’s edge is a duration confound

• At matched duration, CLAP localization beats random removal at every cell on Qwen2-Audio and AF3; blank’s raw edge largely reflects silencing 100% versus 2–4 s.The blank waveform removes a median 10.0 s, while localized windows remove 2–4 s.Source: §5.1 · p. 3

5.3. Oracle-localization control: is the null a localizer artifact?

• Oracle ablation flips 16.8% of claims, versus 10.5% for CLAP’s guess and 3.1% for random, showing both localization error and residual location-insensitivity.The table is an oracle control using synthetic Qwen2-Audio positives with known evidence location and matched-size ablations.Source: Table 2 · p. 3, §5.3 · p. 3

5.2. Graded necessity: thresholding costs the detector

• AUROC=0.651 for continuous final-checkpoint P(yes) drop, versus 0.491 for binary R* on Qwen2-Audio; AF3’s R* is also near chance at 0.52.Adaptive localization efficiency is b∆loc=0.028 and is larger for transient than continuous/ambient classes (0.116 vs. 0.017); mean CLAP–true-location IoU is 0.45, with 25% zero overlap.Source: §5.2 · p. 3

5.4. Language-prior correction: ensemble significant, alone it isn’t

• Language-only prior AUROC=0.490; subtracting it reaches AUROC=0.727, while a three-feature ensemble reaches AUROC=0.743.The ensemble improves detection per-example but not per-class; the language-prior contrast alone does not beat blank-audio detection.Source: §5.4 · p. 3

5.5. Contrastive decoding: a real accuracy gain

• 71.9% ± 0.7pp CV accuracy follows five-paraphrase LangPrior gating with AAD, improving +14.8 over raw and outperforming single-phrasing in 86% of folds.The combined contrast sits above every single contrast across the α range, and confidence-gating raises it further.Source: Figure 1 · p. 4, §5.5 · p. 4, Table 3 · p. 4

5.6. Cross-model generalization: a third model, opposite pattern on AF3

• Qwen2.5-Omni replicates Qwen2-Audio’s raw localized-ablation pattern (0.609 vs. blank’s 0.766), unlike AF3; the three-model evidence tentatively points to encoder lineage over scale.AF3’s localized-versus-random gap is 7–21 points, compared with 0–4 for both Qwen-family models.Source: §5.6 · p. 4, Table 4 · p. 4

5.7. Cross-model transfer of the causal grounding profile

• Cross-model profile transfer underperforms each target model’s best fixed contrast: Qwen→AF3 loses 1.4 pp and AF3→Qwen loses 13.1 pp.The authors interpret the negative transfer as evidence that the models have qualitatively different failure modes: language-prior override versus coarse localization sensitivity.Source: §5.7 · p. 4

6. DISCUSSION, LIMITATIONS, AND CONCLUSION

• The study flags partial CLAP-accuracy effects, per-example-only ensemble significance, no multiple-comparison correction, limited Omni results, and nonfunctional Phi-4/FLAM models.On shorter HalluAudio clips the matched-duration effect attenuates sharply, though a small Qwen2-Audio residual survives; combined decoding replicates cleanly.Source: §6 · p. 4
• Matched-duration tests show localized removal decisively beats a same-size random window on both models, despite model- and class-dependent raw detection value.A single causal grounding profile does not transfer across models.Source: §6 · p. 4

Compliance with Ethical Standards

• The study uses public benchmark audio and pretrained models, collects no new human or animal subject data, and required no institutional ethical review.Source: Compliance with Ethical Standards · p. 4