Source-linked AI summary
Mitigating Cross-Modal Hallucinations by Correcting Modality Misrouting in Omnimodal LLMs
Yu Mi, Licheng Pan, Jiayi Zhu, Qirui Liu, Yitong Qiao, Changtong Xu, Tao Wei, Fei Sun, Zhixuan Chu, Kui Ren
TL;DR
Audio-visual interference can yield incorrect joint predictions even when the question-relevant unimodal branch retains the correct answer. CaRe detects conflicts with a question-selected counterfactual branch and uses confidence-adaptive residual intervention; experiments report consistent improvements over evaluated inference-time baselines across two OLLMs.
Problem
Audio-visual hallucinations can occur when the joint prediction is wrong despite recoverable evidence in the question-relevant unimodal branch.
Source: §1 · p. 1Method
CaRe uses a text-only modality probe and conflict gate to select a counterfactual branch, then applies confidence-adaptive residual intervention to the fused representation while retaining both modalities.
Source: §1 · p. 2Results
CaRe consistently outperforms evaluated inference-time hallucination-mitigation baselines on CMM and AVHBench across two OLLMs.
Source: §1 · pp. 2–3Takeaways & Limitations
The findings support treating some audio-visual hallucinations as correctable when relevant modality-specific evidence remains accessible.
Source: §7 · p. 10CaRe assumes useful task-relevant evidence remains accessible in a modality-specific branch and directly comparable predictions exist in a discrete task-defined answer space.
Source: §7 · p. 10
Abstract
from arXiv · showhide
Omnimodal large language models (OLLMs) are susceptible to cross-modal hallucinations under audio-visual interference. We find that, for a substantial fraction of incorrect joint predictions, the question-relevant unimodal branch can still recover the correct answer, suggesting that task-relevant evidence can remain recoverable from the relevant modality even when it is not reflected in the joint prediction. We refer to this recoverable failure pattern as \emph{modality misrouting}. We propose CaRe (Conflict-aware Residual Intervention), a training-free inference-time method for identifying and correcting modality misrouting. CaRe uses a text-only modality probe to identify a candidate modality-specific counterfactual branch, and flags potential misrouting when its prediction conflicts with the original joint prediction. For such cases, CaRe computes a residual direction from the discrepancy between the counterfactual and fused representations and uses it to steer the original multimodal state toward the relevant modality evidence, while retaining both modalities in the original audio-visual inference path. The intervention strength is adapted based on the confidence of the counterfactual branch. Experiments on CMM and AVHBench across two OLLMs show that CaRe consistently outperforms the evaluated inference-time hallucination-mitigation baselines.
1 Introduction
Audio-visual interference can produce hallucinations even when the question-relevant unimodal branch retains the correct evidence. The introduction presents CaRe as a training-free way to detect and correct this modality misrouting, reporting consistent gains over evaluated inference-time baselines.
1 Introduction
- OLLMs can produce audio-visual responses that are not faithfully grounded in available evidence, limiting reliability in video understanding, question answering, and embodied applications.The figure caption depicts misrouting in both directions and CaRe steering the fused representation toward the selected modality-specific branch.Source: §1 · p. 1, Figure 2 · p. 2
- CaRe consistently outperforms evaluated inference-time hallucination-mitigation baselines across two OLLMs and two audio-visual benchmarks.It retains both modalities in the original multimodal inference path.Source: §1 · pp. 2–3
- CaRe uses a text-only modality probe and a conflict gate to select a counterfactual branch, then steers the fused representation with residual signals.Intervention strength adapts to counterfactual confidence.Source: §1 · p. 2, Figure 1 · p. 2
- Modality misrouting describes incorrect joint predictions for which question-relevant modality-specific evidence remains recoverable but is not reflected in the prediction.The introduction characterizes this as a recoverable failure pattern that can give rise to cross-modal hallucinations.Source: §1 · pp. 1–2
2 Related Work
The related work frames CaRe against omnimodal models whose broad modality coverage does not guarantee reliable evidence integration, and against mitigation methods that have paid less attention to recoverable unimodal evidence.
2.1 Omnimodal Large Language Models
- Omnimodal models unify audio, visual, and textual signals, but competing or imbalanced modalities can keep question-relevant information from being reflected in joint predictions.The section positions CaRe as addressing this inference-time failure without modifying the underlying model.Source: §2.1 · p. 3
2.2 Hallucination Mitigation for Omnimodal Models
- Prior hallucination mitigation uses training-time alignment, representation intervention, or training-free decoding correction, including contrastive decoding, modality reweighting, and activation steering.These approaches are described as primarily suppressing hallucinated outputs or recalibrating modality contributions.Source: §2.2 · p. 3
- The recoverable setting where unimodal evidence remains available but is missed by the joint prediction has received less attention; CaRe detects counterfactual disagreement and selectively corrects the joint representation.CaRe retains the original audio-visual inference path.Source: §2.2 · p. 3
3 Motivation
The motivation section identifies modality misrouting as a recoverable discrepancy between incorrect joint predictions and question-relevant unimodal predictions, then measures how often the unimodal branch recovers the answer.
3 Motivation
- Modality misrouting is the recoverable discrepancy in which the joint branch is wrong while the question-relevant unimodal branch can recover the correct answer.The figure caption characterizes this interference as bidirectional across audio and visual modalities.Source: §3 · p. 3
- 52.9%–74.4% on CMM and 37.3%–46.2% on AVHBench are the target-branch recovery-rate ranges among incorrect joint predictions.The analysis tests whether withholding the competing modality lets the question-relevant branch recover the ground-truth answer.Source: Figure 3 · p. 3, §3 · p. 3
- Recovery across the examined backbones, benchmarks, and interference directions indicates that some joint prediction failures retain usable modality-specific evidence for targeted correction.This motivates correcting the joint representation rather than treating every failure as lost evidence.Source: §3 · pp. 3–4
4 Preliminaries
The preliminaries define joint and unimodal input configurations and use a question-only probe to select a candidate modality-specific reference. Branch predictions are evaluated over task-defined answer choices.
4 Preliminaries
- The method considers joint audio-visual input and audio-only or video-only counterfactual views, formed by withholding the competing modality.The same frozen OLLM runs under each configuration.Source: §4 · p. 4
- Each branch scores answers from the task-defined finite candidate set, and its decision margin indicates how decisively its preferred answer exceeds alternatives.A zero margin means the branch has no unique preferred answer; evaluation uses next-token logits for candidate labels.Source: §4 · p. 4
- The text-only routing probe estimates whether a question needs audio, video, or both, using logits produced with both sensory inputs withheld.It supplies a question-level modality-demand signal without observing the current audio-visual sample.Source: §4 · p. 4
- A modality-specific reference is selected only when the probe uniquely favors audio or video; routing to both or no unique maximum yields no candidate reference.The probe selects a candidate reference but is not assumed to identify the factually correct modality.Source: §4 · p. 4
5 Method
CaRe gates intervention on a decisive conflict between a valid modality-specific reference and the joint prediction, then steers the audio-visual representation toward that reference. It adapts correction strength to counterfactual confidence and decays the intervention across decoder layers.
5.1 Counterfactual Conflict Gate
- The conflict gate intervenes only when a valid routed unimodal reference and both branches make decisive but disagreeing predictions.All other samples retain the original joint prediction.Source: §5.1 · pp. 4–5
5.2 Adaptive Residual Intervention
- For gate-positive cases, CaRe steers the evolving audio-visual state toward the selected counterfactual branch using a layer-wise residual while keeping both modalities available.The residual is computed relative to the current state, including corrections made at preceding layers.Source: §5.2 · p. 5
5.2.1 Confidence-Adaptive Intervention Strength
- CaRe adapts intervention strength to the selected branch’s decision confidence, giving low-confidence predictions conservative corrections and allowing stronger intervention for decisive predictions.The confidence score is derived from the branch’s decision margin; a threshold selects a weak or strong intervention range.Source: §5.2.1 · p. 5
5.2.2 Depth-Decayed Residual Intervention
- Across predefined decoder layers, CaRe decays correction strength from αmax to αmin and modifies only the final prompt-token state during prefilling.Later interventions use the updated audio-visual state, and autoregressive generation proceeds without further intervention on newly generated tokens.Source: §5.2.2 · pp. 5–6
6 Experiments
Experiments compare CaRe with inference-time baselines across two hallucination benchmarks and evaluate its design choices and performance on general audio-visual QA.
6.2 Main Results; 6.4 Performance on General Audio-Visual QA Tasks
- CaRe achieves the best overall accuracy on CMM and AVHBench across both backbones; on Daily-Omni, it modestly raises accuracy from 61.5% to 61.8% and from 61.0% to 61.5%.The main results cover both hallucination benchmarks, while the Daily-Omni comparison assesses general audio-visual QA.Source: §6.2 · pp. 6–7, Table 1 · p. 7, §6.4 · pp. 9–10
6.2 Main Results
- On CMM, CaRe exceeds MAD by 9.8 points on Qwen2.5-Omni-7B and 9.6 points on MiniCPM-o-2.6; on AVHBench, it exceeds MAD by 1.5 and 2.5 points, respectively.On CMM Audio Dominance, accuracy rises from 75.5% to 89.8% for Qwen2.5-Omni-7B and from 72.8% to 88.5% for MiniCPM-o-2.6.Source: §6.2 · p. 7
- On Qwen2.5-Omni-3B, CaRe improves overall accuracy by 5.6 points on CMM and 4.6 points on AVHBench, supporting effectiveness across Qwen model scales.The reported overall accuracies rise from 76.9% to 82.5% on CMM and from 76.4% to 81.0% on AVHBench.Source: §6.2 · pp. 7–8, Table 2 · p. 7
6.3 Ablation Studies
- On gate-positive samples, CaRe outperforms direct counterfactual branch selection by up to 10.82 points on AVHBench and 6.35 points on CMM.The comparison attributes the advantage to residual steering that retains complementary cross-modal context rather than switching directly to the unimodal prediction.Source: §6.3 · p. 8, Table 3 · p. 8
- The default text-only routing prompt achieves 99.5–100% agreement with the task-defined modality across four backbone–benchmark settings, although some prompt and ordering sensitivity remains.Across 96 prompt and option-order configurations, 85 achieve at least 95% agreement.Source: §6.3 · pp. 8–9, Figure 7 · p. 9
- Late intervention layers 25–27 outperform layers 19–21 by 1.6 points on Qwen2.5-Omni-7B and 0.6 points on MiniCPM-o-2.6.The passage reports that early and intermediate ranges provide limited benefits and that accuracy improves sharply in later layers.Source: Figure 4 · p. 8, §6.3 · p. 9
- On MiniCPM-o-2.6, confidence-adaptive strength improves accuracy/F1 by 0.9/0.8 points over all-weak intervention, while depth decay reaches 0.869 accuracy and 0.873 F1.Depth decay also exceeds uniform strength (0.862/0.868) and depth-increasing strength (0.864/0.870).Source: Figure 5 · p. 8, Figure 6 · p. 8, §6.3 · p. 9
7 Conclusion
The conclusion frames some audio-visual hallucinations as recoverable modality misrouting and presents CaRe as a training-free correction. It reports gains on hallucination benchmarks and general audio-visual QA, while noting limits on evidence, task format, and inference cost.
7 Conclusion
- CaRe uses confidence-adaptive, depth-decayed residual intervention at middle-to-late decoder layers without updating model parameters.The conclusion describes detecting conflicts between joint predictions and question-relevant counterfactual views, then using residual signals to correct fused representations.Source: §7 · p. 10
- Experiments show substantial improvements on CMM and AVHBench across two OLLMs, while Daily-Omni performance is preserved and modestly improved.The conclusion characterizes some hallucinations as occurring despite recoverable task-relevant evidence in a modality-specific view.Source: §7 · p. 10
- CaRe assumes relevant evidence remains accessible in a modality-specific branch and uses discrete conflict gating, which may limit it when evidence is weak, conflicts are not explicit, or answers are not readily scorable.The text also notes extra inference overhead from the text-only routing probe and counterfactual branches.Source: §7 · p. 10
AI use statement
- Generative AI assisted with language editing and data organization and formatting, but not methodology, results, or scientific conclusions; authors reviewed the content and take responsibility.Source: AI use statement · p. 10
Reproducibility Statement
- The paper identifies sections and appendices covering the method, evaluation protocol, ablations, implementation, prompts, datasets, and inference costs to support reproduction.Source: Reproducibility Statement · pp. 10–11
A.1 Ablation of the Counterfactual Conflict Gate
The appendix tests CaRe’s conflict gate and the sensitivity of intervention magnitude and confidence allocation, including their joint calibration.
A.1 Ablation of the Counterfactual Conflict Gate
- Removing the conflict condition increases intervention coverage by approximately 75 percentage points on both backbones without changing accuracy or prediction correctness.The comparison indicates that conflict gating excludes redundant corrections and concentrates interventions on samples with disagreement between joint and counterfactual decisions.Source: §A.1 · p. 14, Table 5 · p. 14
A.2 Impact of the Intervention Strength Scale λα
- On CMM, accuracy improves as λα rises from 0.25 to 0.50 and remains stable from 0.50 to 1.25; the default λα = 1 is within 0.2 and 0.1 points of each backbone’s best setting.Larger scales generally produce both more helpful and more harmful prediction changes, indicating a trade-off between correction and over-intervention.Source: §A.2 · pp. 14–15, Figure 8 · p. 15
A.3 Impact of the Confidence Threshold η
- Raising η shifts samples from strong to weak intervention; on Qwen2.5-Omni-7B, weak/strong counts change from 4/163 at η = 0.1 to 138/29 at η = 0.9.The best threshold differs by backbone: Qwen2.5-Omni-7B reaches 86.1% at η = 0.1, while MiniCPM-o-2.6 reaches 86.9% at η = 0.3.Source: §A.3 · pp. 15–16, Figure 9 · p. 15
A.4 Joint Sensitivity to λα and η
- The strength scale and confidence threshold interact: Qwen2.5-Omni-7B holds 86.125% across tested thresholds at λα = 1.25, while MiniCPM-o-2.6 peaks at 86.875% for (λα, η) = (1.0, 0.3).The results support calibrating intervention magnitude and confidence allocation jointly; the default configurations lie in high-performing regions for both backbones.Source: §A.4 · p. 16, Figure 10 · p. 16
B Scaling Analysis
CaRe improves overall performance across Qwen2.5-Omni-3B and 7B while retaining gains across cross-modal hallucination directions. Its inference algorithm activates confidence-scaled residual correction only after a modality probe selects a candidate branch whose prediction conflicts with the joint output.
B Scaling Analysis
- CaRe consistently improves overall performance across Qwen2.5-Omni-3B and 7B, with gains across different cross-modal hallucination directions.Both models use the same evaluation protocol.Source: B Scaling Analysis · p. 17
C.1 CaRe Inference Algorithm
- CaRe probes question-level modality demand, runs a counterfactual branch only when audio or video is uniquely favored, and corrects only hard conflicts.Otherwise, it retains the original audio-visual output.Source: §C.1 · p. 17
- For gate-positive cases, CaRe applies layer-wise residuals to the final prompt token, scaling intervention strength by counterfactual confidence and decreasing it with depth.Generation then continues on the original audio-visual path without further intervention.Source: §C.1 · p. 17, §C.2 · p. 18
C.2 Hyperparameter Configurations
- Qwen2.5-Omni-7B and MiniCPM-o-2.6 use layers 25–27, while Qwen2.5-Omni-3B uses layers 33–35; the paper also examines hyperparameter sensitivity.These are the final three language-decoder layers for each backbone.Source: §C.2 · pp. 17–18
C.3 Baseline Implementation Details
The evaluation uses official implementations for several inference-time baselines and adapts VCD to omnimodal inputs by contrasting the full input with modality-withheld branches. The text-only modality probe uses a default prompt and tests alternative formulations and option orderings for robustness.
C.3 Baseline Implementation Details
- ICT, AVCD, and MAD use their official open-source implementations, interfaced directly with the omnimodal backbones.Source: §C.3 · p. 18
- The VCD adaptation contrasts complete audio-visual input with unimodal and text-only sub-branches, withholding competing sensory inputs rather than adding heuristic noise.The original VCD was designed for vision-language models and contrasted original representations with visual perturbations.Source: §C.3 · pp. 18–19
D Prompt Formulations for the Text-Only Modality Probe
- Main experiments use prompt D with options ordered audio, video, both; robustness analysis tests alternative prompt formulations and option orderings.Prompts N1–N3 share the same modality-routing objective as prompt D.Source: D Prompt Formulations for the… · p. 19, Table 8 · p. 19
E Additional Dataset Details
The additional dataset details describe two benchmarks focused on cross-modal hallucination and one broader audio-visual question-answering benchmark. The selected tasks test bidirectional audio-visual interference as well as general audio-visual understanding.
E.1 CMM
- CMM evaluation uses Visual Dominance and Audio Dominance, with 200 instances and 400 binary Yes–No questions per subset, for 800 questions total.The subsets examine visual content inducing absent sounds and auditory cues inducing hallucinated visual content.Source: §E.1 · p. 19
E.2 AVHBench
- AVHBench evaluation uses Audio-driven Video Hallucination (1,136 questions) and Video-driven Audio Hallucination (2,290 questions), both balanced binary Yes–No tasks.The tasks test hallucinations induced across modalities in opposite directions.Source: §E.2 · p. 19
E.3 Daily-Omni
- Daily-Omni contributes 1,197 multiple-choice QA pairs across six categories and 30- or 60-second clips to assess broader audio-visual understanding.The full benchmark is used beyond the hallucination-focused settings of CMM and AVHBench.Source: §E.3 · p. 20
F Computational Analysis
CaRe adds latency relative to AVCD but is faster than MAD in serial execution, while much of its runtime comes from the original and counterfactual inference branches. Ideal branch-level parallel execution is estimated to reduce CaRe latency by about 29%.
End-to-end latency
- AVCD has the lowest mean latency on both backbones; CaRe reduces MAD latency by 59.7% on Qwen2.5-Omni-7B (6.70 vs 16.61 s) and 41.5% on MiniCPM-o-2.6 (4.57 vs 7.82 s).Source: F Computational Analysis · p. 20
Runtime decomposition of CaRe
- The original audio-visual and counterfactual branches account for most runtime, while the probe and gating add negligible overhead; residual correction costs 0.305 s and 0.040 s, respectively.The costs are for Qwen2.5-Omni-7B and MiniCPM-o-2.6.Source: F Computational Analysis · p. 20, Table 10 · p. 21
Potential for parallel execution
- Ideal branch-level parallel execution estimates CaRe latency falling from 6.70 to 4.78 s on Qwen2.5-Omni-7B and 4.57 to 3.24 s on MiniCPM-o-2.6.These correspond to estimated reductions of 28.6% and 29.2%, respectively.Source: F Computational Analysis · p. 21, Table 11 · p. 21