AVES-DPO: Self-Corrected Preference Learning Reduces Hallucinations in LVLMs

Researchers propose AVES-DPO (Alignment via VErified Self-correction DPO), a framework to mitigate hallucinations in Large Vision-Language Models (LVLMs). Unlike existing preference learning methods that rely on proprietary models—causing distributional mismatch—AVES-DPO uses the model's own intrinsic knowledge to generate in-distribution preference pairs. A consensus-based verification mechanism diagnoses diverse hallucinations and guides the model to self-correct. Experiments show AVES-DPO surpasses baselines in hallucination mitigation using only 5.2k samples. The work is published on arXiv.