CFPO: Counterfactual Policy Optimization for Multimodal Reasoning

The paper introduces CounterFactual Policy Optimization (CFPO), a novel framework designed to enhance multimodal reasoning in Large Vision-Language Models (LVLMs) by incorporating counterfactual learning mechanisms. CFPO employs a cross-modal counterfactual enhancement approach that regularizes the policy to improve causal consistency between visual and textual inputs, yielding performance improvements of 3.17%-6.25% over standard reinforcement learning baselines and 1.32%-2.13% over the state-of-the-art perception-aware method. This advancement is significant for practitioners as it addresses grounding failures and hallucination issues in LVLMs, promoting more reliable and accurate reasoning in AI applications.