Adversarial Orthogonal Disentanglement (AOD) mitigates hallucinations in large vision-language models (LVLMs)

The researchers propose the Adversarial Orthogonal Disentanglement (AOD) framework, which learns hallucination-related directions with a minimax objective to mitigate LVLM hallucinations. This direction supports training-free dual-forward contrastive decoding, suppressing hallucinations while preserving general capabilities. On four hallucination and four utility benchmarks across three LVLMs, POPE accuracy improves by an average of over 6%, AMBER improves by 6%, utility tasks such as MMMU remain strong, and cross-dataset transfer is robust.