Groc-PO: Grounded Context Preference Optimization for Real-World Multimodal Large Language Models

Groc-PO mitigates truthfulness issues such as visual hallucination and content confabulation in multimodal large language models through staged grounded preference supervision. Addressing the shortcoming that standard DPO typically optimizes only the final answer and early grounding errors easily propagate across stages, it builds a multi-stage preference dataset, GCPD, around three stages: object grounding, context grounding, and grounded reasoning. Experiments show that Groc-PO outperforms standard DPO and strong baselines in hallucination mitigation, faithful reasoning, and overall reliability.