DyCo-RL: Dynamic Cross-Modal Coordination for Visual Reasoning

Why do visual reasoning failures persist even after RLVR training ?

We find that reasoning failures are often associated with breakdowns in this coordination process.
We find that the issue is often not visual perception error or text reasoning error alone, but a failure of dynamic cross-modal coordination. During Chain-of-Thought generation, successful reasoning requires models to continuously switch between looking at visual evidence and thinking on previously established textual context. Existing RLVR methods optimize final outcomes but largely ignore this token-level behavior.

Through token-level analyses and causal interventions, we show that reasoning failures frequently occur when visually-oriented tokens stop attending to relevant image content, or when text-oriented tokens fail to remain grounded in prior reasoning history.

To address this problem, we introduce DyCo-RL, a plug-and-play RLVR framework that explicitly rewards effective cross-modal coordination. DyCo-RL identifies token functional roles using Fisher–Rao attention dynamics and reweights policy optimization according to role-attention alignment. The resulting models exhibit substantially stronger reasoning performance across diverse visual and mathematical reasoning benchmarks.