Research map / Large language models / Instruction and fine-tuning
Differential Fine-Tuning Large Language Models Towards Better Diverse Reasoning Abilities
Xiaosong Yuan, Chen Shen, Shaotian Yan, kaiyuan liu, Xiaofeng Zhang, Sinan Fan, Liang Xie, Wenxiao Wang, Renchu Guan, Ying Wang, Jieping Ye
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DiFT to Mistral-7B mix-up Math+Code SFT | Mix-Math-Code · Mistral-7B ATA | ATA: 52.87 -> 54.80 | Improved |
| add DiFT to continual Math-Code SFT | Continual-Math-Code · Mistral-7B ATA | ATA: 64.58 -> 65.81 | Improved |
| add DiFT to continual Math-Logic SFT on Mistral-7B | Continual-Math-Logic · Mistral-7B ATA | ATA: 16.94 -> 20.06 | Improved |
| add DiFT to Code+Logic mix-up on Llama3-8B | Mix-Code-Logic · Llama3-8B GSM8k | accuracy: 52.31 -> 41.09 | Hurt |
| add DiFT to Instruct model Logic+CSQA mix SFT | Llama3-8B-Instruct · LogiQA2 | accuracy: 32.64 -> 34.48 | Improved |
| fine-tune only non-DSR parameters (inverse DiFT) in mix-up (ablation) | DiFT · Mix-Math-Code GSM8k | accuracy: 61.26 vs 67.02 | Improved |
| fine-tune only DSR in continual SFT (inverse DiFT) (ablation) | DiFT · Continual-Math-Code GSM8k | accuracy: 25.17 vs 46.32 | Improved |
| use top-DSR rows instead of random rows of same count, mix-up (ablation) | DiFT · Mix-Math-Code GSM8k | accuracy: 62.35 -> 67.02 | Improved |
| use top-DSR rows instead of random rows, continual (ablation) | DiFT · Continual-Math-Code xGLUE | pass rate: 0.8192 -> 1.0557 | Improved |
| increase DSR rows from 100 to 200 (ablation) | DiFT · GSM8k | accuracy | Hurt |
| transplant CSQA DSR into math-tuned model (ablation) | math-sft · GSM8k | accuracy: 61.64 -> 41.02 | Hurt |
| transplant math DSR into CSQA-tuned model (ablation) | csqa-sft · GSM8k | accuracy: 8.79 -> 22.21 | Improved |
Closest papers by meaning
- UFT: Unifying Supervised and Reinforcement Fine-Tuning (NeurIPS 2025 · similarity 0.92)
- The First Few Tokens Are All You Need: An Efficient and Effective Unsupervised Prefix Fine-Tuning Method for Reasoning Models (NeurIPS 2025 · similarity 0.91)
- Training Language Models to Reason Efficiently (NeurIPS 2025 · similarity 0.91)
- TAIA: Large Language Models are Out-of-Distribution Data Learners (NeurIPS 2024 · similarity 0.91)
- Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.90)
- QFFT, Question-Free Fine-Tuning for Adaptive Reasoning (NeurIPS 2025 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
