A manuscript marked under review presents MetaCtrl as a lightweight controller for regulating a frozen reasoning model. Its stated objective is to decide when additional reasoning is useful while balancing correctness against shorter reasoning trajectories.
The controller observes the evolving reasoning trace and selects among four text-based interventions; continue, fast think, skip think, and stop think. These interventions are applied at sentence-ending paragraph boundaries; stop think prompts the reasoner to conclude rather than forcibly truncating generation.
Across the reported evaluations, MetaCtrl improved or maintained accuracy while reducing generated-token length on the seen DeepSeek-R1-Distill-Qwen-7B reasoner and on unseen reasoners including Qwen3-8B, Qwen3-14B, and DeepSeek-R1-Distill-Qwen-32B.
The reported experiments also found lower mean end-to-end inference latency than Vanilla Reasoning across evaluated separate-GPU and shared-GPU configurations. The controller adds memory requirements and system complexity, so the measured speedups do not by themselves establish end-to-end service economics.
Transfer tests used a controller trained with a frozen DeepSeek-R1-Distill-Qwen-7B reasoner and evaluated it without further training on other reasoners, including a different Llama-based architecture, and across additional benchmarks.