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

TL;DR

This paper introduces RC2, a reinforcement learning framework that enforces cycle consistency across modalities to improve multimodal reasoning accuracy without relying on labeled data.

Contribution

The paper proposes a novel cycle-consistent reinforcement learning approach that aligns internal representations across modalities, enhancing reasoning performance.

Findings

Improves reasoning accuracy by up to 7.6 points.

Enables autonomous alignment of internal representations.

Reduces modality-specific errors.

Abstract

Robust perception and reasoning require consistency across sensory modalities. Yet current multimodal models often violate this principle, yielding contradictory predictions for visual and textual representations of the same concept. Rather than masking these failures with standard voting mechanisms, which can amplify systematic biases, we show that cross-modal inconsistency provides a rich and natural signal for learning. We introduce RC2, a reinforcement learning framework that resolves internal conflicts by enforcing cross-modal cycle consistency. By requiring a model to perform backward inference, switch modalities, and reliably reconstruct the answer through forward inference, we obtain a dense, label-free reward. This cyclic constraint encourages the model to align its internal representations autonomously. Optimizing for this structure mitigates modality-specific errors and improves reasoning accuracy by up to 7.6 points. Our results suggest that advanced reasoning emerges not only from scaling data, but also from enforcing a structurally consistent understanding of the world.