Groc-PO introduces grounded context preference optimization to reduce multimodal hallucinations

Researchers propose Grounded Context Preference Optimization (Groc-PO), a framework designed to address untruthfulness issues like visual hallucinations and content fabrication in Multimodal Large Language Models (MLLMs). The method tackles the credit-assignment challenge where standard Direct Preference Optimization (DPO) fails to suppress error propagation from early grounding stages by applying supervision only at the final answer.

• Groc-PO introduces explicit preference supervision across multiple grounded stages rather than just the final output.
• The authors construct the Grounded Context Preference Dataset (GCPD) with multi-stage samples covering Object Grounding, Contextual Grounding, and Grounded Reasoning.
• This approach aims to capture the formation, integration, and utilization of grounded context to mitigate cross-stage error propagation.

Extensive experiments demonstrate that Groc-PO improves performance in hallucination mitigation, faithful reasoning, and overall reliability compared to standard DPO and other baselines.