R-C2: Breaking the Bottleneck of Multimodal Reasoning with Cross-Modal Cycle-Consistent Reinforcement Learning

Rutgers University and other institutions proposed the R-C2 framework, which converts cross-modal inconsistencies in multimodal models into self-supervised learning signals. Through cycle consistency constraints, it achieves improved reasoning capabilities without manual annotation, gaining up to 7.6 percentage points in performance across multiple benchmark tests, providing a new path to address the "modality gap" dilemma in multimodal reasoning.

Current Multimodal Large Language Models (MLLMs) face the "modality gap" problem: different modal inputs of the same content may lead to contradictory answers. Traditional solutions like large-scale fine-tuning rely on expensive manual annotation and are difficult to scale; reinforcement learning lacks reliable reward signals; majority voting mechanisms tend to reinforce systemic biases and cannot resolve inter-modal or intra-modal inconsistencies.

The core of the R-C2 framework is a "forward-reverse-reconstruction" cycle verification process: given a candidate answer, the model performs reverse reasoning to generate a query, then switches modalities and performs forward reasoning to reconstruct the original answer. This cycle forms four-way cross-validation (T→T, T→I, I→T, I→I), using cycle consistency as an unlabeled reward signal to drive the model to optimize cross-modal representation alignment without manual annotation of question-answer pairs.

The research team validated the effectiveness of R-C2 on multiple authoritative benchmarks such as ScienceQA, ChartQA, and MathVista, achieving up to a 7.6 percentage point improvement in reasoning accuracy on models with 3B and 8B parameters. It also significantly improved cross-modal prediction consistency, and the higher the task modality complexity (e.g., MathVista), the more obvious the gains.

R-C2 proposes a new perspective on AI development: advanced reasoning capabilities do not only come from expanding data scale but also require enforcing the structural consistency of the world. This framework represents the ability of "self-supervised metacognition", where the model actively checks the consistency of its own reasoning, providing key insights for building autonomous and reliable AI systems.

R-C2 has limitations such as high computational cost and difficulty in achieving consistent representations for extremely challenging samples. Future directions include expanding to more modalities, exploring efficient cycle verification strategies, and combining with supervised fine-tuning to form a hybrid training paradigm.