MixReasoning: Switching Modes to Think

Reasoning models enhance performance by tackling problems in a step-by-step manner, decomposing them into sub-problems and exploring long chains of thought before producing an answer. However, applying extended reasoning to every step introduces substantial redundancy, as sub-problems vary widely in difficulty and complexity: a small number of pivotal steps are genuinely challenging and decisive for the final answer, while many others only involve straightforward revisions or simple computations. Therefore, a natural idea is to endow reasoning models with the ability to adaptively respond to this variation, rather than treating all steps with the same level of elaboration. To this end, we propose MixReasoning, a framework that dynamically adjusts the depth of reasoning within a single response. MixReasoning enables fine-grained mode switching by training a lightweight concise LoRA adapter and control its strength to trigger switches based on reasoning difficulty estimated from sliding-window token confidence, yielding human-like transitions between fast and slow reasoning. The resulting chain of thought then becomes a mixture of detailed reasoning on difficult steps and concise inference on simpler ones. Experiments on AIME24, MATH-500, GPQA, and GSM8K demonstrate that MixReasoning shortens reasoning length by 13\%--49\% across benchmarks of varying difficulty, delivering consistent efficiency gains while maintaining performance.

Large reasoning models often solve problems by generating many intermediate steps before giving an answer. This helps them handle complex tasks, but it also creates a problem: not every step needs the same amount of thought. Some parts of a solution are genuinely difficult, such as deciding the plan or making a key derivation, while many others are routine, such as simple calculations or straightforward follow-up steps. Treating all steps as equally difficult can make reasoning models unnecessarily slow, costly, and verbose. This paper introduces MixReasoning, a framework that allows a reasoning model to dynamically adjust the reasoning depth within a single response. When the model encounters a difficult or uncertain part, it uses more detailed reasoning; when the next part is easier, it switches to a more concise style. MixReasoning achieves this by adding a lightweight adapter for concise reasoning and using the model’s own token-level confidence to estimate which parts of the response need deeper thought. Across several math and science reasoning benchmarks, MixReasoning reduces the length of generated reasoning by 13%–49% while maintaining comparable performance. These results suggest that reasoning models can be made more efficient by allocating their effort more selectively: thinking carefully where it matters, and staying concise where it does not.