You can steer LLM reasoning in real-time by treating it as a control problem: a separate agent learns to guide the main model's thinking steps, saving tokens while maintaining accuracy and letting you trade off speed vs. quality.
This paper introduces ACTS, a method that uses a controller agent to guide how a language model reasons during inference. Instead of letting the model think freely, the controller observes the reasoning progress and remaining token budget, then suggests what strategy to use next—enabling efficient reasoning with explicit control over the thinking process.