Omni-Woodpecker: Correcting Omni-modal Large Language Model Hallucinations via Evidence-Prior Rebalancing

Omni-Woodpecker is a training-free framework that reduces cross-modal hallucinations in omni-modal large language models by identifying cross-modal conflicts, separating evidence from misleading priors via orthogonal decomposition, and rebalancing fused representations before decoding. With Qwen2.5-Omni-7B as the backbone, it improves CMM and AVHBench by 12.9 and 8.6 points respectively, and generalizes to VideoLLaMA2-AV (improvements of 4.6 and 3.9 points).