Control-R: Towards controllable test-time scaling

The paper introduces Reasoning Control Fields (RCF), a novel test-time method that injects structured control signals to guide and adjust reasoning effort in Large Reasoning Models, addressing issues of underthinking and overthinking in long chain-of-thought reasoning. It also presents the Control-R-4K dataset with annotated reasoning processes and proposes a Conditional Distillation Finetuning (CDF) approach, achieving state-of-the-art results on benchmarks like AIME2024 and MATH500 with controllable reasoning capabilities