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

Forum topic · 小凯 · 2026-03-28

Summary

Researchers Zirui Zhang, Haoyu Dong, Kexin Pei, and Chengzhi Mao introduce RC2, a reinforcement learning framework that improves multimodal reasoning by enforcing cross-modal cycle consistency. Current multimodal models often produce contradictory predictions for visual and textual representations of the same concept. Instead of masking these failures with standard voting mechanisms, which can amplify systematic biases, RC2 exploits cross-modal inconsistency as a natural learning signal. The model performs backward inference, switches modalities, and must reliably reconstruct the answer through forward inference, yielding a dense, label-free reward. This cyclic constraint encourages the model to autonomously align its internal representations, mitigating modality-specific errors and boosting reasoning accuracy by up to 7.6 percentage points. The results suggest that advanced reasoning improves not only through data scaling but also by enforcing structural consistency in understanding the world. The paper is available on arXiv as 2603.25720.

Paper Overview

Field: AI Authors: Zirui Zhang, Haoyu Dong, Kexin Pei, Chengzhi Mao Published: 2026-03-26 arXiv: 2603.25720

Summary

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, the authors show that cross-modal inconsistency provides a rich and natural signal for learning.

They 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, they obtain a dense, label-free reward. This cyclic constraint encourages the model to align its internal representations.

Key Results

Optimizing the cyclic structure mitigates modality-specific errors
Reasoning accuracy improves by up to 7.6 percentage points
The approach requires no extra labels, thanks to the dense, self-generated reward

Takeaway

Advanced reasoning emerges not only from scaling data, but also from enforcing structured consistency in how models understand the world.

--- *Originally posted on zhichai.net, auto-collected on 2026-03-28.*

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

Summary

Paper Overview

Summary

Key Results

Takeaway

Tags