Recent LLM-based control systems typically handle decision-making in one of two ways: they either call a language model to select each action, or they synthesize a world model and plan against it at every step. Both approaches introduce latency that can limit real-time performance.
The paper introduces Code to Control, a method that instead synthesizes parameterized reactive controllers. By generating executable code for the controller, the approach avoids the need for per-decision LLM invocations or online planning. The abstract suggests this design reduces the computational overhead that has constrained LLM-based control in real-time settings.
The work is described as a way to address the latency bottleneck, though the abstract does not provide experimental validation. The contribution is the controller synthesis approach itself, which shifts the focus from runtime inference to generating reactive policies in code.