Researchers have built a chart-to-code AI system that deliberately keeps its visual and coding skills apart, rather than training them together.
The method, called MoCA (Mixture of Cross-modal Arbitration), tackles a specific problem: models that turn chart images into executable code usually learn to "see" and to "code" either in isolation or fused into one process. Both approaches ignore that the two skills work differently and can interfere with each other during training. MoCA instead runs a separate visual branch and a separate code branch, with a lightweight "arbiter" component deciding how much each branch contributes at every layer and every generated token. Training happens in two stages: a supervised warm-up using self-distilled reasoning steps, then reinforcement learning that rewards both the reasoning process and the final code output. On three benchmarks, MoCA performs competitively against both general-purpose models and chart-specialized ones.
The interesting part isn't the benchmark scores, which the researchers themselves frame as "competitive" rather than dominant. It's the ablation finding that gains come from splitting the two abilities and letting the arbiter allocate credit dynamically, not from simply adding parameters. That's a useful data point for anyone building multimodal systems, since the industry's default move is usually to throw more scale at entangled training rather than rethink the architecture.
Chart-to-code is a narrow task, but it's a clean testbed for a broader question: when should a model's skills be kept separate, and when should they merge? MoCA's answer, for now, is that separation wins when the two tasks reason about fundamentally different things.