R-C2: Cycle-Consistent Reinforcement Learning Improves Multimodal Reasoning Robustness

arXiv·medium signal

Zhang, Dong, and Pei propose R-C2, a reinforcement learning approach that enforces cycle consistency across modalities — if a model reasons about an image, converts to text, then back to image, the answer should be consistent. Current multimodal models often produce contradictory predictions for visual and textual representations of the same concept. R-C2 uses this consistency constraint as a reward signal, producing more robust multimodal reasoning without additional data or human annotation.