acceptodds
Withdrawn from ICLR 2027

From Attention Weights to Audio Writes: Mitigating Hallucinations in Large Audio Language Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large audio language models (LALMs) may describe sound events or attributes that are absent from the input recording, resulting in hallucinations. Existing attention interventions mitigate these errors by sustaining attention to audio, but attention weights alone do not determine the vector contributions of audio positions to the attention output. To address this limitation, we introduce the Audio Relative Write Ratio (ARWR), which measures the norm of the vector update from audio positions relative to the sum of update norms from audio and other positions. By incorporating value aggregation and output projection, ARWR provides a more accurate characterization of the contribution of audio positions to the attention output. Building on this diagnostic, we propose ARWR-Cal, a framework for hallucination mitigation through residual calibration of the final hidden state. At each generation step, a lightweight low-rank calibrator uses the source vectors and final hidden state to predict a residual adjustment. ARWR controls the permitted adjustment magnitude. Experiments show ARWR outperforms audio attention in hallucination diagnosis, while ARWR-Cal improves average accuracy by 8.2 percentage points over the strongest baseline across three datasets and three LALMs. Code is available at https://anonymous.4open.science/r/ARWR-Cal.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Audio Language Models, Hallucination Mitigation, Residual Calibration

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_5C3wOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes ARWR, which measures the relative strength of audio writes to the attention output using vector norms after value aggregation and output projection, instead of attention weights alone. ARWR-Cal uses this ratio to cap a residual correction to the final hidden state, predicted by a small calibrator on top of a frozen backbone. The method is evaluated on three backbones and three datasets, and ARWR is also tested as a hallucination-risk signal on Clotho-AQA.

Critical strengths

S1. Gains across datasets and backbones. ARWR-Cal has the best accuracy and F1 on the Adversarial and Popular conditions of AudioHallucination for all three backbones (Table 1). On HalluAudio it wins all nine backbone–domain combinations, by 11.7–27.3 points over the strongest baseline in each. It is also best on Clotho-AQA for every backbone, though the margins there are small. For me, the HalluAudio numbers are the strongest part of the paper.

S2. The diagnostic claim is checked with numbers. ARWR accounts for value aggregation and output projection, so it can pick up contributions that attention weights miss, and the paper checks whether this actually matters. On Qwen2-Audio answers from Clotho-AQA, ARWR gets an AUROC of 0.7431 vs. 0.6295 for audio attention mass, and the 95% CI of the paired difference stays above zero (§5.4, Table 3). The two extra samples in Appendix C.2 agree.

Critical weaknesses

W1. The smaller gains under Random are not explained. I put a lot of weight on this condition. Popular and Adversarial target errors driven by object frequency and co-occurrence, while Random checks whether the model can judge presence or absence from the audio when negatives are not picked by either criterion. ARWR-Cal allows larger corrections when ARWR is low, so I want to know how its effect changes across these question types.

In Table 1, ARWR-Cal improves Random accuracy over Base by 8.1–14.1 points but is still 9.2–9.6 points behind AHA, the best baseline there. Its gain over Base is smallest under Random on all three backbones. Most other methods do best under Random. ARWR-Cal does worst there on every backbone. The authors mention the dependence on negative sampling (§5.2) and the trade-off between correction and preserving correct answers (§5.3). Per-condition accuracy and F1 are reported, and Appendix D has a few Clotho-AQA examples of fixed and broken answers. What I could not find is a paired before/after analysis for each AudioHallucination condition. Without it, I can't tell whether the Random gap comes from fixing fewer errors, breaking more correct answers, or gains concentrated in some question types. That makes it hard to judge how far the method's benefit extends.

W2. The ablations mix up ARWR-based budget allocation with correction size. In Eq. 6 the adaptive budget lies in bt∈[0.5,1.5]b_t\in[0.5,1.5], but the fixed-budget control (§5.3, Table 2) uses the lowest value, bt=0.5b_t=0.5. Since the budget directly scales the correction in Eq. 10, this comparison changes both input-dependent adaptation and the size of the correction at once. The other two ablations don't resolve this. w/o Source Features keeps the ARWR budget, and w/o Adaptive Budget keeps the source features. What is missing is a calibrator that uses only final hidden-state features with a fixed budget, trained with the same supervision. Without it, it is hard to see what ARWR and the source features add over plain calibration.

This matters for Random too. On Qwen2-Audio, the fixed-budget variant gets 77.4% Random accuracy vs. 70.0% for the full model (Table 2). SALMONN-7B doesn't show this (65.0% for the full model vs. 64.1% for the fixed budget, Appendix Table 5), so I'm not saying adaptive budgets hurt Random in general. I just think the effect of the budget on correction and preservation should be checked across backbones.

I see real value in the cross-dataset gains and in ARWR as a diagnostic. But the Random pattern and the added value of the ARWR budget over simpler calibration aren't explained yet, so for now I lean toward weak reject.

Questions for authors

Q1. The smaller gains over Base under Random are the main reason I'm not more positive about this paper. What drives this pattern? I'd like the authors to look into it with targeted analyses, plus additional experiments if needed. Some places to start: how predictions change after calibration, how results vary by question and sound type, and whether the calibration-data composition or the budget settings matter. These are just suggestions. Please test the explanations you find most plausible and report what you find, including results that don't support them, any trade-offs, and what is still unclear. It would also help to keep what the evidence shows separate from what is still a hypothesis. If this is addressed well, I'd be quite willing to raise my score.

Q2. How are the three AudioHallucination conditions used in calibrator training, and do ARWR-Cal and AHA train on the same data? Since there is one calibrator per backbone–dataset pair, I'd like to know the number of calibration-pool samples per condition, whether any weighting is used, and how duplicate questions across conditions are handled. I'm also not sure whether AHA's "existing training partitions" contain the same audio and questions as ARWR-Cal's training data. Actual training sample counts for both methods would clear this up.

Q3. Does the ARWR-guided budget still help against a well-chosen fixed budget, e.g. one tuned on the validation set or matched to the adaptive budget's mean? A fixed-budget calibrator using only final hidden-state features, trained on the same data with the same objective, would also show what ARWR and the source features bring.

Q4. Do ARWR and the applied correction size behave differently under Random? The ARWR distribution and the actual correction magnitudes per condition, linked to corrected errors and to correct answers that flip to wrong, would help locate the issue. On a separate note, Appendix B.3 counts only false Yes answers as hallucinations. Among answers predicted as Yes, can ARWR separate correct from incorrect ones? That would show whether it tracks correctness or just answer polarity.

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

The paper argues that audio attention mass is an incomplete indicator of audio's contribution to decoding in large audio language models (LALMs), because it ignores value vectors and the output projection (Kobayashi et al., 2020). It proposes the Audio Relative Write Ratio (ARWR): the norm of the attention-output contribution from audio positions, relative to the sum of audio and non-audio contribution norms, averaged over layers [D/2, 3D/4). On top of ARWR, ARWR-Cal trains a small residual MLP (k = 8) on the final hidden state, taking source vectors and the hidden state as input, on a 10% in-domain calibration pool per model–dataset pair; ARWR sets the permitted residual norm via a budget b_t ∈ [0.5, 1.5].

Evidence comes in two parts. (i) Diagnosis: on 1,000 Clotho-AQA answers from Qwen2-Audio, ARWR detects false-affirmative hallucinations better than attention mass (AUROC 0.743 vs 0.630; paired 95% CI of the difference [0.082, 0.145]). (ii) Mitigation: on three benchmarks and three LALMs, ARWR-Cal improves average accuracy by 8.2 pp over AHA (reproduced from Table 1 as the dataset-level mean), supported by two single-component ablations.

The diagnostic analysis is careful, but the central claim — that ARWR guides effective mitigation — is not separated from the effect of supervised in-domain calibration, and the diagnostic evidence covers a single setting. I would raise my score if W1 is resolved in favor of ARWR.

Critical strengths

S1. Well-controlled diagnostic comparison. Both signals are scored on the same answers with the same labels. Hallucination rate rises monotonically across ARWR quintiles (4.5% → 43.0%) but not across attention-mass quintiles (it falls from 30.5% in group 4 to 27.0% in group 5); the paired AUROC CI excludes zero, and the trend survives two re-samplings (App. C.2). This is solid evidence that information beyond attention weights matters.

S2. Write vectors carry usable signal, and gains are consistent. Adding source features improves average accuracy by 5.5 pp (Qwen2-Audio) and 7.1 pp (SALMONN-7B) (Tables 2, 5). ARWR-Cal beats all baselines on HalluAudio for all three backbones while training only a tiny module on a frozen backbone.

Critical weaknesses

W1. ARWR's contribution to the mitigation gains is not isolated (primary concern). (a) No "neither component" baseline. Removing one component keeps much of the margin over AHA: 45% / 31% without source features and 70% / 33% without the adaptive budget (Qwen2-Audio / SALMONN-7B, average accuracy). Without a hidden-state-only calibrator under a constant budget, the share due to plain supervised calibration is unknown. (b) The budget ablation is confounded with capacity. Since τ is the 75th percentile of ARWR, about 75% of training samples get b_t > 0.5, whereas the control fixes b = 0.5, the minimum. The gain may thus partly reflect a larger average budget rather than adaptivity. (c) Diagnosis and gains do not align. ARWR is validated only on Clotho-AQA, where ARWR-Cal gains just 1.2–2.1 pp over the best baseline; the largest gains (12–27 pp) are on HalluAudio, where no ARWR–hallucination link is shown. (d) No mechanism check. The paper does not show that ‖Δh_t‖ or answer flips concentrate at low ARWR. Needed: a constant budget matched to mean b_t; shuffled-ARWR budgets; entropy- or attention-driven budgets; a plain hidden-state calibrator; an SFT baseline on the same pool; and ‖Δh_t‖ and flip rate as functions of ARWR.

W2. The diagnostic evidence is narrow. (a) One model, one dataset, one error type (false "yes"). App. C.2 re-samples the same pool, so it is not an independent replication; yet the abstract claims superiority in hallucination diagnosis generally. (b) Weak comparison set. Only raw attention mass is compared. Missing: sink-removed attention mass, ‖u_A‖ alone, P(yes) or entropy, and a hidden-state probe. (c) Possible question-type confound. Since h = 1 only when the model answers "yes" and all references are "no" (App. B.3), all reference-"yes" questions are negatives; ARWR may partly track whether the queried sound is present. Please report AUROC within reference-"no" questions.

Questions for authors

see weakness

Reviewer_6oiaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper targets hallucination in large audio language models , i.e., answers that assert sound events or attributes absent from the recording. It argues that attention-based interventions such as AttnAdapter, which sustain attention mass on audio tokens, are mis-targeted because attention weights alone do not determine the vector contribution of audio positions . Two contributions are proposed:

  1. ARWR (Audio Relative Write Ratio), a diagnostic defined as ∥uA∥2/(∥uA∥2+∥uN∥2)\|u^A\|_2 / (\|u^A\|_2 + \|u^N\|_2), where uAu^A and uNu^N are the audio and non-audio parts of the attention output.
  2. ARWR-Cal, a mitigation method in which a small trained calibrator predicts a residual adjustment Δht\Delta h_t to the final hidden state. Its norm is bounded by an ARWR-dependent budget (Eq. 6–7, 10), and the backbone stays frozen.

ARWR outperforms audio attention mass in hallucination detection (AUROC: 0.743 vs. 0.630), while ARWR-Cal improves average accuracy by 8.2 percentage points over the baseline AHA.

Critical strengths

Moving from attention weights to the post-WOW^O vector write is a sound refinement grounded in established interpretability work, and it is cheap to compute. The evidence is the most convincing part of the paper. Under ARWR ranking the hallucination rate rises monotonically from 4.5% to 43.0% across risk groups, whereas under attention mass it peaks at 30.5% in the fourth group and falls to 27.0% in the fifth group.

Critical weaknesses
  1. Base models are outdated.
  2. On HalluAudio speech, every baseline stays at 32–43% accuracy, and the re-trained AHA barely moves from Base (Qwen2-Audio 37.3 vs 34.6; SALMONN-13B 42.7 vs 42.8). Yet the same AHA gains up to about 25 points on AudioHallucination. AHA is trained with rank-16 LoRA at learning rate 2×10−62\times10^{-6}, which is very low for LoRA. It is also unclear how open-ended and multiple-choice HalluAudio answers are extracted and graded. Both factors may favor a method that edits logits directly.
  3. A flat mean over the seven accuracy columns of Table 1 gives about 9.9 points. Nearly all of it comes from HalluAudio. The paper discloses the random-subset loss but offers no explanation.
Questions for authors

N/A

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

Related papers

Discussion (0)

Sign in to comment.