MoCA: A New Approach to Chart-to-Code Generation with Cross-modal Arbitration

A new method called MoCA (Mixture of Cross-modal Arbitration) has been proposed for chart-to-code generation, which involves reading fine-grained visual details of a chart and writing executable code that reproduces it. Existing methods either train visual and coding abilities separately or fine-tune on chart-to-code data with the two abilities entangled, neither accounting for the distinct nature of the abilities or interference when optimized together. MoCA separates the two abilities rather than blending them, built on a Cross-modal Arbitration Block (CAB) that maintains a visual branch and a code branch as distinct pathways, with a lightweight arbiter arbitrating their relative contributions at every layer and generated token. Training occurs in two stages: supervised warm-up on self-distilled reasoning trajectories that decompose visual understanding into explicit steps, followed by reinforcement learning. The paper is available on arXiv with identifier 2608.15510.