PaperlayerView exact revision on arXiv

Source-linked AI summary

Relevant Evidence Decoding for Audio-Visual Hallucination Mitigation

Hyunjae Ra, Aecheon Jung, Jungin Park, Sungeun Hong

arXiv:2610.02976v1cs.AI

TL;DR

Joint audio-visual inference can weaken predictions despite sufficient evidence in the question-required modality, and existing approaches do not align corrections with each question’s evidence needs. RED identifies required evidence and selectively strengthens its PMI-measured contribution; across AV-LLMs and benchmarks, it generally outperforms existing decoding methods.

  • Problem

    Joint audio-visual inference can weaken predictions even when the question-required modality alone provides sufficient evidence for the correct answer.

    Source: §1 · pp. 1–2

  • Method

    RED uses a question-only pass to select required evidence and PMI to selectively strengthen its predictive contribution relative to the question and language context.

    Source: §1 · p. 2, §3.3 · p. 5

  • Results

    Across three AV-LLMs and five benchmarks, RED generally outperforms existing methods on hallucination and general audio-visual reasoning, with average relative TTFT of 1.5× standard decoding.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Controlled interventions show increased sensitivity to question-required evidence and reduced sensitivity to irrelevant evidence with RED.

    Source: §1 · p. 2, §5 · p. 9

    The method is evaluated for autoregressive AV-LLMs with frozen parameters and makes modifications only at decoding time.

    Source: §3.1 · p. 3

Abstract

from arXiv · show

Audio-Visual Large Language Models (AV-LLMs) remain prone to cross-modal hallucinations, where one modality incorrectly affects predictions about another. Although contrastive decoding reduces hallucinations in vision-language models, its direct extension to AV-LLMs overlooks a key challenge: different questions require different perceptual evidence, including audio, video, or their interaction. Notably, we observe that joint audio-visual inference can weaken the prediction even when a model can recover the correct answer from a single informative modality. For example, when asked which instrument is heard, a model may correctly predict violin from the audio alone. Once a video showing a guitar is added, its confidence in violin may drop. In this paper, we introduce Relevant Evidence Decoding (RED), a training-free method that identifies question-relevant evidence and selectively strengthens its contribution. RED uses pointwise mutual information to quantify the predictive support provided by audio and video beyond the question alone. It decomposes their joint contribution into audio, video, and residual interaction components. A question-only inference pass determines the required evidence type, after which the model augments the original audio-visual prediction with the corresponding PMI contribution. Across three audio-visual hallucination benchmarks and three AV-LLMs, RED improves accuracy over standard decoding by up to 7% on CMM, 6.3% on AVHBench, and 3.8% on SVHalluc, with an average relative time to first token of 1.5x standard decoding.

1 INTRODUCTION

Joint audio-visual inference can weaken predictions even when the question-required modality supports the correct answer, motivating decoding that follows question-relevant evidence. RED selects that evidence and its predictive support; the paper reports gains across audio-visual hallucination and reasoning benchmarks.

1 INTRODUCTION

  • Joint audio-visual inference can weaken a prediction despite sufficient evidence in the question-required modality, while correct answers can rely on correlated or irrelevant cues.The proposed evidence-following criterion calls for changing predictions when required evidence changes and reducing their response to changes in non-required evidence.Source: §1 · pp. 1–2, Figure 1 · p. 2
  • Prior AV-LLM decoding methods balance or integrate modalities, but do not align the correction with the evidence required by each question.These approaches include modality-dominance correction, adaptive modality-specific branches, and cross-modal information-flow modeling.Source: §1 · p. 1
  • Contrastive decoding adjusts predictions using differences across conditioning inputs without additional training, but correct answers alone do not establish appropriate evidence use.The introduction frames RED’s correction around question-required predictive support rather than any signal that favors an answer.Source: §1 · pp. 1–2
  • RED uses a question-only pass to identify required evidence, then uses PMI to measure and selectively strengthen its audio, video, or residual interaction contribution.PMI measures predictive support beyond the question and language context.Source: §1 · p. 2
  • Across three AV-LLMs and five benchmarks, RED generally outperforms existing decoding methods on hallucination and general audio-visual reasoning tasks.The evaluation also reports transfer to two general audio-visual question-answering benchmarks.Source: §1 · p. 2

2 RELATED WORK

Related work covers multimodal hallucination and decoding-based mitigation, while distinguishing question-conditioned evidence following from input dependence or modality balancing. RED selects contributions defined relative to a shared question-only reference.

Multimodal hallucination

  • Audio-visual hallucinations have been linked to modality dominance, cross-modal interference, and spurious inter-modality correlations, while input dependence alone does not show whether evidence required by a question is followed.Evidence following is framed as responding to required evidence while remaining stable to non-required evidence.Source: §2 · p. 3

Decoding-based hallucination mitigation

  • AVCD, MAD, and ASD use modality-dependent contrasts, adaptive branches, and cross-modal information flow, respectively.MAD is described as closest to RED: it adaptively weights branches, whereas RED selects audio, video, or residual contributions defined relative to a common question-only reference.Source: §2 · p. 3

3 METHOD

RED uses question-guided evidence selection and PMI-based contrasts to adjust AV-LLM predictions according to the perceptual evidence a query requires. Its motivation is that joint audio-video inputs can weaken reliance on required evidence while leaving predictions sensitive to irrelevant evidence.

3.1 PRELIMINARIES

  • Contrastive decoding adjusts the original prediction using differences between input conditions, but its meaning depends on how the reference condition is constructed.The method describes the model as frozen and modifications as decoding-time only.Source: §3.1 · p. 3

3.2 MOTIVATION

  • 59.7% sensitivity with only required evidence fell to 44.0% with both modalities, while non-required-modality interventions changed 17.0% of predictions.The comparison motivates correcting predictive influence rather than only final prediction performance.Source: Figure 3 · p. 4, §3.2 · p. 5

3.3 RELEVANT EVIDENCE DECODING

  • PMI measures each evidence contribution as a prediction change from the same question-only reference, separating audio, video, joint, and residual interaction effects.The interaction contribution is the non-additive residual not explained by separate audio and video contributions.Source: §3.3 · p. 5
  • RED makes a question-only pass to select none, audio, video, audio plus video, or audio-visual interaction as the required evidence type.For two-modality questions, audio plus video means both are needed; interaction denotes an explicitly queried relation such as correspondence or temporal alignment.Source: §3.3 · p. 5
  • RED selects the contribution matching the chosen evidence type and uses it to augment the original audio-visual prediction during decoding.It evaluates only auxiliary question-only, audio, and video branches required by the selection; other decoding operations remain unchanged.Source: Figure 2 · p. 4, §3.3 · p. 6

4 EXPERIMENTS

RED is evaluated across hallucination and general audio-visual QA benchmarks, with analyses examining discrimination, component contributions, evidence selection, and inference cost.

4.2 MAIN RESULTS

  • On CMM, RED raises Qwen2.5-Omni-7B accuracy from 82.2% to 89.2%; on AVHBench, it beats the strongest competing methods by 2.1–3.0% across backbones.CMM gains concentrate on visual and audio dominance subsets, at 16.5% and 17.3%, respectively.Source: §4.2 · p. 6
  • RED achieves the best SVHalluc accuracy across three backbones, raising standard decoding from 53.3–60.5% to 57.1–63.4%.SVHalluc requires grounding semantically rich speech with visual evidence.Source: §4.2 · p. 7
  • On MUSIC-AVQA-Hard, RED raises Qwen2.5-Omni-7B accuracy from 31.9% to 40.5%, exceeding MAD by 4.5%.On MUSIC-AVQA, RED improves accuracy from 64.5% to 67.1%, matching MAD.Source: §4.2 · p. 7

4.3 ANALYSES

  • Across the three-backbone mean, RED increases TPR from 78.8% to 84.7% and lowers FPR from 23.6% to 19.3%; the Base threshold sweep cannot match its operating point.The analysis separates perceptual grounding gains from shifts in response bias.Source: §4.3 · pp. 7–8, Figure 4 · p. 8
  • On AVHBench, adding PMI decomposition to Q-guided selection raises accuracy from 76.2% to 78.6% and AUROC from 82.7% to 86.1%.The ablation finds both question-guided selection and PMI decomposition necessary for the full RED formulation.Source: §4.3 · p. 8
  • Selected target-only substantially improves over Base, but RED further improves both MUSIC-AVQA benchmarks; mismatched and random evidence assignments underperform RED.Competitive Heuristic and Oracle results support the importance of aligning corrections with question-required evidence, rather than a particular selector implementation.Source: §4.3 · p. 8
  • RED raises mean accuracy from 71.0% to 75.3% across nine settings at 1.5× TTFT, and improves over MAD from 73.6% to 75.3% while reducing TTFT from 2.6× to 1.5×.Figure 5 reports accuracy and relative TTFT averaged across nine backbone-benchmark settings.Source: §4.3 · pp. 8–9, Figure 5 · p. 9
  • On CMM, RED yields lower NLL than three reference settings when audio and video contributions are scaled with evidence selection fixed.Qualitative examples show RED answering cases on MUSIC-AVQA-Hard where MAD fails.Source: §4.3 · p. 9, Figure 7 · p. 9, Figure 6 · p. 9

5 CONCLUSION

The conclusion frames cross-modal hallucination as an evidence-following problem: predictions can be correct yet weakly depend on the perceptual evidence a question requires. RED selectively augments audio-visual predictions using question-only evidence selection and corresponding centered contributions; the appendix reports centering behavior and robustness.

5 CONCLUSION

  • RED decomposes audio-visual predictive changes relative to a question-only reference and selectively augments the contribution matching the question-required evidence.The conclusion reports that controlled interventions increase sensitivity to required evidence and reduce sensitivity to irrelevant evidence.Source: §5 · p. 9
  • As a training-free method, RED requires no task-specific retraining and may reduce computational burden and additional use of potentially sensitive audio-visual data.The authors suggest these properties may facilitate use in resource-constrained and privacy-sensitive settings.Source: ETHICS STATEMENT · p. 10

A.1 QUESTION-GUIDED EVIDENCE SELECTION

  • The frozen backbone selects among five evidence types from question text alone, without audio or video input, and that selection determines which contribution is added.The types are None, Audio, Video, Both, and Relation.Source: §A.1 · p. 14

A.2 EFFECT OF SCORE CENTERING

  • Increasing the unlabeled centering sample budget does not improve performance monotonically, and the accuracy benefit of centering persists across five draws for Qwen2.5-Omni-7B.The other backbones show less consistent benefits, with greater sampling variability for VideoLLaMA2.1-7B-AV.Source: §A.2 · p. 14, VideoLLaMA2.1-7B-AV · p. 15
  • On CMM, centering raises Qwen2.5-Omni-7B accuracy from 87.3% to 89.2%, while AUROC changes only marginally across all three backbones.The reported AUROC pairs are 95.4% to 95.5%, 93.6% to 93.6%, and 92.5% to 92.8%; the authors interpret this as mainly correcting systematic score offsets at the final decision stage.Source: §A.2 · pp. 14–15

A.3 DIAGNOSTICS OF EVIDENCE CONTRIBUTION SELECTION

The diagnostics test whether decoding benefits depend on selecting the appropriate evidence contribution, rather than merely adding more modality information. Results favor contributions matched to the selected evidence type, while mismatched corrections move the operating point in the opposite direction.

A.3 DIAGNOSTICS OF EVIDENCE CONTRIBUTION SELECTION

  • The diagnostic mapping assigns audio, video, both-modality, and relation evidence to ∆A, ∆V, ∆A + ∆V, and residual ∆I, respectively.The reported correction terms use centered contributions.Source: §A.3 · p. 15

Contribution choice matters

  • Across eligible groups, RED moves toward higher TPR and TNR, whereas mismatched corrections move in the opposite direction.Adding both unimodal contributions offers limited benefit compared with RED, indicating that more contributions are not necessarily better than a contribution matched to the question's evidence requirement.Source: §A.3 · p. 17, Figure D · p. 17

Loss landscape on CMM

  • The loss-landscape diagnostic varies audio and video contribution weights while holding evidence selections, predictions, and centering offsets fixed.RED uses fixed coefficients: (1, 0) for Audio and (0, 1) for Video; the sweep is diagnostic and does not tune decoding strengths.Source: §A.3 · p. 17

Modality-specific and residual contributions

  • For AVHBench, the best contribution matches the evidence requirement: audio ∆A for VDA, video ∆V for ADV, and residual ∆I for AVM.Table B compares accuracy and AUROC with Qwen2.5-Omni-3B while holding centering offsets fixed.Source: §A.3 · p. 18

Acc. AUROC Acc. AUROC Acc. AUROC

On MUSIC-AVQA-Hard, additive audio and video contributions perform best for questions tagged as requiring both modalities, but the prompt-selected subset favors video alone. The results distinguish contribution choice from whether question-only selection identifies the required evidence.

Additive contributions for both-modality evidence

  • 40.16% accuracy: on the Oracle-tagged subset, ∆A + ∆V outperforms either modality-specific contribution alone.The analysis uses MUSIC-AVQA-Hard because A + V cases are too sparse on AVHBench for a meaningful direct analysis.Source: Table C · p. 18, Acc. AUROC Acc. AUROC Acc. AUROC · p. 18
  • 38.63% vs. 38.19%: on the Prompt-selected subset, ∆V slightly outperforms ∆A + ∆V.Prompt-selected assignments need not coincide with dataset Audio-Visual evidence tags, so effectiveness depends on identifying the required evidence as well as mapping it to a contribution.Source: Table C · p. 18, Acc. AUROC Acc. AUROC Acc. AUROC · p. 18

Evidence selection across benchmarks

  • Selector predictions vary across benchmarks: CMM predominantly selects Audio or Video, while Relation is substantial on AVHBench and Both is most frequent on MUSIC-AVQA and MUSIC-AVQA-Hard.These are question-only predicted evidence requirements, not ground-truth modality dependencies.Source: Acc. AUROC Acc. AUROC Acc. AUROC · pp. 18–19

A.4 INFERENCE EFFICIENCY ANALYSIS

RED adds inference cost by evaluating auxiliary prediction paths, but selecting evidence incurs little overhead and avoids unnecessary perceptual processing. Its reported time to first token remains below MAD and ASD across the nine benchmark-backbone settings, and a CMM comparison reports a better accuracy–TTFT trade-off than aggressive SID pruning.

A.4 INFERENCE EFFICIENCY ANALYSIS

B.1 EVALUATION PROTOCOL

  • RED's evaluation uses frozen backbones and deterministic decoding on complete CMM, AVHBench, and SVHalluc benchmarks, with unweighted task means.For score statistics, RED samples 100 unlabeled questions per benchmark and uses the resulting statistics throughout evaluation.Source: §B.1 · p. 20

B.2 BASELINES

  • The comparison covers audio-visual hallucination methods and vision-language hallucination methods adapted to audio, video, and joint inputs while retaining their decoding rules.ASD's attention intervention is adapted to binary question answering.Source: §B.2 · p. 20

B.3 BENCHMARK DETAILS AND METRICS

The evaluation spans three audio-visual hallucination benchmarks and two general audio-visual question-answering datasets, using accuracy as the primary metric. Each benchmark reports averages over its defined tasks or subsets.

CMM

  • CMM measures hallucinations across six subsets covering spurious cross-modal correlations and visual, audio, or language dominance, with each subset containing 200 samples.The complete benchmark has 1,200 samples and 2,400 Yes/No questions; results report the unweighted mean across subsets.Source: §B.3 · p. 21

AVHBench

  • AVHBench evaluates audio-driven video hallucination, video-driven audio hallucination, and audio-visual matching across 5,302 questions over 2,136 videos.The reported accuracy is the unweighted mean over its three judgment tasks.Source: §B.3 · p. 21

SVHalluc

  • SVHalluc covers three semantic and three temporal speech–vision tasks using 2,405 video-question pairs from YouCook2’s validation split.Results include unweighted semantic and temporal means and an unweighted average across all six tasks; evaluation is zero-shot.Source: §B.3 · pp. 21–22

MUSIC-AVQA

  • MUSIC-AVQA evaluates open-ended musical-performance questions across audio, visual, and joint audio-visual settings; this study uses 9,129 corrected test questions.The reported metric is overall accuracy on the complete corrected test split.Source: §B.3 · p. 22

MUSIC-AVQA-Hard

  • MUSIC-AVQA-Hard is an audio-sensitive subset that removes questions answered correctly in two central-frame-only GPT-4o runs, leaving 2,508 questions for evaluation.Overall accuracy is reported on the filtered subset.Source: §B.3 · p. 22

Metrics

  • Accuracy is the primary metric; AUROC is additionally reported for analyses of continuous binary answer scores to assess discrimination independently of a fixed threshold.CMM, AVHBench, SVHalluc, and both MUSIC-AVQA splits use their specified unweighted task or subset means where applicable.Source: §B.3 · p. 22

B.4 EVIDENCE-INTERVENTION PROTOCOL AND RESULTS

The intervention protocol tests whether predictions track answer-changing swaps of question-required evidence while remaining stable under answer-preserving swaps of non-required evidence. RED improves sensitivity and reduces invariance violations across audio- and video-required questions.

Intervention construction and evaluation

  • The evaluation uses 3,359 retained AVHBench intervention pairs: 1,621 answer-changing sensitivity pairs and 1,738 answer-preserving invariance pairs.These pairs come from 2,680 distinct source video-question items.Source: §B.4 · pp. 22–23

Annotation-guided pair construction

  • For sensitivity pairs, donors have explicit opposite-label annotations for the same queried entity and required modality; the question and other modality are retained during replacement.For invariance pairs, the required stream is retained while the non-required stream is replaced, leaving the annotated answer unchanged.Source: §B.4 · pp. 22–23

Annotation and media verification

  • Media verification checks source provenance, answer transitions, and evidence consistency, while AVHBench annotations determine reference answers and methods share retained pairs and donor assignments.Audio and sampled video evidence are also examined with PANNs and OWLv2, respectively.Source: §B.4 · p. 23

Pair-level evaluation

  • Sensitivity requires correctly answering both members of an answer-changing pair, while invariance violation is the fraction of answer-preserving pairs whose predictions change.The Target-only reference has zero invariance violation by construction because swapping a non-required modality leaves its input unchanged.Source: §B.4 · p. 23

Evidence following under intervention

  • Audio-required sensitivity rises from 28.0% for Base to 58.3% with RED, while invariance violation falls from 26.3% to 9.2%; video-required sensitivity rises from 60.0% to 70.1%, with violation falling from 7.7% to 3.9%.Across methods using both modalities, RED has the highest sensitivity for both evidence types and the lowest audio-required invariance violation; MAD is slightly lower than RED on video-required invariance violation (3.5% vs. 3.9%).Source: §B.4 · pp. 23–24

C.1 DETAILED CMM AND SVHALLUC RESULTS

Across CMM and SVHalluc, RED achieves the highest average accuracy across all three backbones, with CMM gains concentrated on modality-dominance subsets. Additional analyses report evidence recovery under full audio-visual input, a comparison with PMI-based methods, and MUSIC-AVQA accuracies.

C.1 DETAILED CMM AND SVHALLUC RESULTS

  • RED achieves the highest average accuracy across all three backbones on both CMM and SVHalluc.The detailed results are reported by subset for CMM and by task for SVHalluc.Source: §C.1 · pp. 24–25, Table I · p. 26
  • On CMM, RED is best or tied best on Vis. and Aud. across all three backbones; Qwen2.5-Omni-7B gains 16.5% on Vis. and 17.3% on Aud. over Base.Changes on VL, AL, and VAL are smaller.Source: §C.1 · pp. 24–25
  • SVHalluc gains are uneven: CMSB improves over Base for all three backbones, while temporal changes depend more on the backbone.VideoLLaMA2.1-7B-AV’s near-chance GSA and TA results match a reported benchmark failure pattern; the benchmark attributes GSA behavior to positive-correspondence alignment bias.Source: §C.1 · p. 24, Table I · p. 26

C.2 EVIDENCE ATTENUATION AND RECOVERY

  • On AVHBench, the required modality substantially raises correct-answer probability over question-only inference, but full audio-visual input attenuates part of that gain; RED recovers and further strengthens support.The averages cover audio-required and video-required questions across three backbones.Source: §C.2 · p. 25, §C.4 · p. 27

C.3 COMPARISON WITH PMI-BASED DECODING

  • On CMM with Qwen2.5-Omni-3B, RED achieves the highest average accuracy among the compared PMI-based decoding methods.Table J reports accuracy, with higher values better.Source: §C.3 · pp. 25–26

C.4 DETAILED MUSIC-AVQA RESULTS

  • RED reaches 67.1% overall accuracy on the 9,129-question corrected MUSIC-AVQA test split and 40.5% on the 2,508-question MUSIC-AVQA-Hard subset.Tables K and L provide the corresponding question-type-level results.Source: §C.4 · pp. 26–27

D QUALITATIVE ANALYSIS

The qualitative analysis presents individual AVHBench examples of evidence attenuation and representative hallucinations. In both, RED restores or corrects predictions while retaining the perceptual streams.

D.1 EVIDENCE ATTENUATION AND RECOVERY

  • Across one audio-required and two video-required examples, the required modality alone strongly supports the correct answer, while RED restores and strengthens that support with both streams present.These are individual examples illustrating the evidence-attenuation pattern.Source: §D.1 · p. 27

D.2 REPRESENTATIVE HALLUCINATION CASES

  • Representative audio, visual, and alignment hallucination cases show RED correcting predictions inconsistent with the question-required evidence while retaining both perceptual streams.The examples use Qwen2.5-Omni-3B.Source: §D.2 · p. 27

E LIMITATIONS

The evaluation is limited to audio-visual question answering, and RED's evidence mapping has boundaries for finer-grained questions and other modality configurations.

E LIMITATIONS

  • The evaluation covers binary, multiple-choice, and open-ended audio-visual question answering, but not long-form tasks such as audio-visual captioning.Source: E LIMITATIONS · p. 27
  • RED uses five discrete audio-visual evidence types, but finer-grained questions or other modality configurations may require a different evidence space and selection mechanism.Source: E LIMITATIONS · p. 28
  • Extending RED beyond audio-visual settings may require redefining the contribution decomposition for the modalities involved.Source: E LIMITATIONS · p. 28

Related papers

Loading 2610.02976v1…