arxiv:2506.12860

QFFT, Question-Free Fine-Tuning for Adaptive Reasoning

Published on Jun 15

· Submitted by

lwl-uestc on Jun 18

Upvote

Authors:

Wanlong Liu ,

Abstract

Question-Free Fine-Tuning (QFFT) improves efficiency and adaptability in cognitive models by leveraging both short and long chain-of-thought patterns, reducing response length while maintaining performance across various scenarios.

AI-generated summary

Recent advancements in Long Chain-of-Thought (CoT) reasoning models have improved performance on complex tasks, but they suffer from overthinking, which generates redundant reasoning steps, especially for simple questions. This paper revisits the reasoning patterns of Long and Short CoT models, observing that the Short CoT patterns offer concise reasoning efficiently, while the Long CoT patterns excel in challenging scenarios where the Short CoT patterns struggle. To enable models to leverage both patterns, we propose Question-Free Fine-Tuning (QFFT), a fine-tuning approach that removes the input question during training and learns exclusively from Long CoT responses. This approach enables the model to adaptively employ both reasoning patterns: it prioritizes the Short CoT patterns and activates the Long CoT patterns only when necessary. Experiments on various mathematical datasets demonstrate that QFFT reduces average response length by more than 50\%, while achieving performance comparable to Supervised Fine-Tuning (SFT). Additionally, QFFT exhibits superior performance compared to SFT in noisy, out-of-domain, and low-resource scenarios.

View arXiv page View PDF GitHub repository Add to collection

Community

lwl-uestc

Paper author Paper submitter 4 days ago

Models fine-tuned with Long CoT data often overthink simple questions, generating long-winded reasoning. A new SFT approach, QFFT, tackles this by removing questions during training and using only Long CoT answers for fine-tuning.

This method preserves the model’s native Short CoT ability, avoiding the “question→Long CoT” mapping trap of traditional SFT. By learning from answer structures, models can activate Long CoT’s deep thinking when facing errors or complex problems, while defaulting to efficient Short CoT for easy tasks.

The results are impressive: a 40% reduction in response length on math tasks, comparable performance to SFT, and superior noise resistance and cross-domain generalization. In noisy data, QFFT maintains 78.6% performance versus SFT’s 0.4%, and it excels in low-resource scenarios.