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 R-C2, 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 voting mechanisms, which can amplify systematic biases, R-C2 exploits cross-modal inconsistency as a natural training signal. The model must perform backward inference, switch modalities, and 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, reduces modality-specific errors, and boosts reasoning accuracy by up to 7.6 percentage points. The results suggest that advanced reasoning comes not only from scaling data but also from enforcing structural consistency in how models understand the world. Paper available on arXiv: 2603.25720.

Paper Overview

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

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

They introduce R-C2, 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, the approach obtains a dense, label-free reward. This cyclic constraint encourages the model to align its internal representations.

Key Findings

Optimizing this structure mitigates modality-specific errors and improves reasoning accuracy by up to 7.6 percentage points.
Cycle consistency offers a label-free reward signal, avoiding the cost of annotated supervision.
The results suggest that advanced reasoning emerges not only from scaling data, but also from enforcing structural consistency in a model's understanding of the world.

--- *Auto-collected on 2026-03-28*