Mitigating Cross-Modal Hallucinations by Correcting Modality Misrouting in Omnimodal LLMs

Key points

• Omnimodal large language models (OLLMs) are susceptible to cross-modal hallucinations under audio-visual interference.1
• 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.1
• 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.1
• Experiments on CMM and AVHBench across two OLLMs show that CaRe consistently outperforms the evaluated inference-time hallucination-mitigation baselines.1

Sources (1)

• [1]Mitigating Cross-Modal Hallucinations by Correcting Modality Misrouting in Omnimodal LLMs
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 10:48 AM

  We propose CaRe (Conflict-aware Residual Intervention), a training-free inference-time method for identifying and correcting modality misrouting.

  Omnimodal large language models (OLLMs) are susceptible to cross-modal hallucinations under audio-visual interference.

Extractive summary: sentences quoted from the sources.