Research map / Large language models / Instruction and fine-tuning
Difficulty Is Not Enough: Curriculum Learning for LLMs Fine-tuning Must Consider Utility
Zishang Jiang, Jinyi Han, Tingyun Li, Xinyi Wang, Sihang Jiang, Xiaojun Meng, Jiansheng Wei, Jiaqing Liang, Yanghua Xiao
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 |
|---|---|---|---|
| use DUCL curriculum instead of random order | LLaMA3-8B CPT · Proof-Pile-2 downstream average | accuracy: 31.68 -> 33.57 | Improved |
| use DUCL curriculum instead of PPL difficulty-only order | LLaMA3-8B CPT · Proof-Pile-2 downstream average | accuracy: 32.32 -> 33.57 | Improved |
| use DUCL instead of PPL order | LLaMA3-8B CPT · MMLU-EM | accuracy: 37.30 -> 40.36 | Improved |
| use DUCL instead of PPL order | LLaMA3-8B CPT · GSM8K | accuracy: 24.87 -> 25.55 | Improved |
| use PPL difficulty-only curriculum instead of random order | LLaMA3-8B CPT · validation loss convergence | steps to equal loss: 2600 steps vs 3000 steps for RANDOM; DUCL 2000 | Hurt |
| use DUCL instead of random/PPL for faster convergence | LLaMA3-8B CPT · validation loss | training steps: 3000 -> 2000 steps | Improved |
| use DUCL instead of random order | Qwen2.5-7B-Instruct SFT · MATH500 | accuracy: 80.4 -> 81.6 | Improved |
| use DUCL instead of random order | Qwen2.5-7B-Instruct SFT · AIME24 | accuracy: 23.3 -> 26.7 | Improved |
| use DUCL instead of PPL order | Qwen2.5-7B-Instruct SFT · AIME25 | accuracy: 16.7 -> 23.3 | Improved |
| use DUCL instead of random order | Qwen2.5-7B-Instruct SFT · IFEval (forgetting) | accuracy: 54.6 -> 57.7 | Improved |
| use DUCL instead of random order | Qwen2.5-7B-Instruct SFT · average over 7 benchmarks | accuracy: 53.8 -> 55.2 | Improved |
| use DUE scoring with strict ordering (alpha=1.0) instead of PPL ordering (ablation) | DUCL · CPT early stage | Val. loss: 1.506 -> 1.474 | Improved |
Closest papers by meaning
- Unveiling the Secret Recipe: A Guide For Supervised Fine-Tuning Small LLMs (ICLR 2025 · similarity 0.91)
- Long Is More for Alignment: A Simple but Tough-to-Beat Baseline for Instruction Fine-Tuning (ICML 2024 · similarity 0.90)
- Get more for less: Principled Data Selection for Warming Up Fine-Tuning in LLMs (ICLR 2024 · similarity 0.89)
- Two-stage LLM Fine-tuning with Less Specialization and More Generalization (ICLR 2024 · similarity 0.88)
- Fixing It in Post: A Comparative Study of LLM Post-Training Data Quality and Model Performance (NeurIPS 2025 · similarity 0.88)
- EvaLearn: Quantifying the Learning Capability and Efficiency of LLMs via Sequential Problem Solving (NeurIPS 2025 · similarity 0.88)
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.
