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

arXiv cs.AI / 3/27/2026

💬 OpinionSignals & Early TrendsIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper argues that multimodal models often produce contradictory outputs across modalities (e.g., text vs. vision) and that these inconsistencies can be leveraged as a learning signal rather than hidden with voting.
It introduces RC2, a reinforcement learning framework that enforces cross-modal cycle consistency by performing backward inference, switching modalities, and then forward-reconstructing the answer.
RC2 uses the cyclic reconstruction objective to generate a dense, label-free reward signal that encourages alignment of internal representations.
Experiments reportedly improve multimodal reasoning accuracy by up to 7.6 points, and the authors suggest gains come from structurally consistent world understanding in addition to scaling.

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.