Recent Omni large language models (OLLMs) have made strides in joint audio-visual reasoning, but the arXiv paper notes that cross-modal hallucinations remain a major challenge. These errors occur when the model generates content inconsistent with the combined audio and visual evidence.
The paper, titled Omni-Woodpecker, points out that existing mitigation approaches mainly operate at the output level. In contrast, the proposed method is described as correcting hallucinations via evidence-prior rebalancing. The abstract does not provide further detail on the mechanism in the available excerpt.
Because the source is a single arXiv announcement, the summary reflects only that one description. No independent comparison or evaluation of Omni-Woodpecker is included in the provided material.