The paper introduces a framework for evaluator-guided test-time reasoning, where a model generates answers and an evaluator scores them. Rather than treating this as a single pass, the authors model each step as a local control problem: should the system trust the current answer, fall back to an alternative, or spend more compute on further inference? This is formalized as a three-action controller that is explicitly compute-aware, meaning the cost of additional reasoning is part of the decision.

The key contribution is the framing itself. By making the trade-off between accuracy and inference budget explicit at every step, the controller can be calibrated against confidence scores. This allows the system to decide when to stop expanding reasoning, potentially reducing wasted computation while maintaining or improving answer quality.

The abstract is brief and does not include experimental results or comparisons to other methods, so the practical gains are not yet established. Still, the formulation offers a clean way to think about the growing problem of balancing reasoning depth with compute limits in large language models.