Wanlong Liu Junxiao Xu Fei Yu Yukang Lin Ke Ji Wenyu Chen Yan Xu Yasheng Wang Lifeng Shang Benyou Wang

Abstract

Recent advancements in Long Chain-of-Thought (CoT) reasoning models haveimproved performance on complex tasks, but they suffer from overthinking, whichgenerates redundant reasoning steps, especially for simple questions. Thispaper revisits the reasoning patterns of Long and Short CoT models, observingthat the Short CoT patterns offer concise reasoning efficiently, while the LongCoT patterns excel in challenging scenarios where the Short CoT patternsstruggle. To enable models to leverage both patterns, we propose Question-FreeFine-Tuning (QFFT), a fine-tuning approach that removes the input questionduring training and learns exclusively from Long CoT responses. This approachenables the model to adaptively employ both reasoning patterns: it prioritizesthe Short CoT patterns and activates the Long CoT patterns only when necessary.Experiments on various mathematical datasets demonstrate that QFFT reducesaverage response length by more than 50\%, while achieving performancecomparable to Supervised Fine-Tuning (SFT). Additionally, QFFT exhibitssuperior performance compared to SFT in noisy, out-of-domain, and low-resourcescenarios.