Research map / Large language models / Instruction and fine-tuning

QFFT, Question-Free Fine-Tuning for Adaptive Reasoning

NeurIPS 2025 · Large language models › Instruction and fine-tuning · approach: Fine-tuning for reasoning data · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Question-Free Fine-Tuning (QFFT), which trains on Long CoT responses with the question removed, so Qwen2.5-Instruct models default to Short CoT and invoke reflective Long CoT only when needed, cutting tokens by about 50% at SFT-level accuracy.

Problem

Long CoT reasoning models overthink simple questions, producing redundant tokens, while standard SFT on (question, Long CoT) pairs overrides the model's default Short CoT patterns and makes it over-rely on Long CoT.

Approach

Remove the input question from SFT data and compute the loss only over the Long CoT response, so the model does not learn a fixed Q-to-Long-CoT mapping; it keeps its Short CoT default and learns reflective behaviors that transfer to Short CoT contexts when uncertainty arises.

Key results

• QFFT reduces average tokens by about 50% (e.g., 7B S1.1: 8.2K to 5.3K average; GSM8K 1.7K to 0.4K) with average accuracy within about 1 point of SFT across GSM8K, MATH, and AIME25.
• On noisy data (Level IV irrelevant answers), SFT drops from 76.5% to 0.4% on MATH while QFFT keeps 78.6%.
• In the 7B long-to-short comparison, QFFT achieves AES 3.4 average versus 2.4 for DAD-7B and -3.6 for SimPO-Shortest.

Measured effects reported in the paper

Surprising findings

• Training only on Long CoT responses without questions makes the model default to Short CoT, with over 50% Short CoT responses even without questions, and only 0.1% to 1% question-paired samples are enough to override this.
• Reflective Long CoT behaviors transfer to Short CoT contexts: Long CoT models showed reflection in 71.2% of cases after erroneous Short CoT steps versus 42.6% after correct ones.

Limitations

• QFFT does not optimize Long CoT efficiency, so overthinking persists on hard questions such as AIME24 and AIME25.
• Evaluation is mostly on math datasets; out-of-domain results are limited to GPQA, MMLU-Pro, and LLM-AggreFact.
• The theoretical equivalences (null-question SFT, continued pre-training) and transfer claim are argued rather than rigorously proven.

Experimental setup

Datasets: S1.1 LIMO Bespoke-Stratos-17k GSM8K ProcessBench

Benchmarks: GSM8K MATH500 AMC23 Minerva AIME24 AIME25 GPQA MMLU-Pro LLM-AggreFact

Metrics: accuracy average tokens RAK (reasoning adaptability Cohen's kappa) AES (accuracy-efficiency score)

Baselines: SFT Difficulty-Adaptive Distillation (DAD) SFT-Shortest DPO-Shortest SimPO-Shortest O1-Pruner

Builds on: DeepSeek-R1 distillation S1 LIMO O1-Pruner SimPO

Closest papers by meaning

• Ada-R1: Hybrid-CoT via Bi-Level Adaptive Reasoning Optimization (NeurIPS 2025 · similarity 0.90)
• SuCo: Sufficiency-guided Continuous Adaptive Reasoning (ICML 2026 · similarity 0.89)
• SCOUT: Teaching Pre-trained Language Models to Enhance Reasoning via Flow Chain-of-Thought (NeurIPS 2025 · similarity 0.89)
• Facts Don't Speak Louder than Words: The Behavioral Essence of Long-CoT Distillation (NeurIPS 2026 · similarity 0.89)
• Expanding Reasoning Potential in Foundation Model by Learning Diverse Chains of Thought Patterns (ICLR 2026 · similarity 0.89)
• Demystifying Long Chain-of-Thought Reasoning (ICML 2025 · similarity 0.89)