AI Paper Atlas

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

AAAI 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning generalization and delta compression · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
use DUCL curriculum instead of random orderLLaMA3-8B CPT · Proof-Pile-2 downstream averageaccuracy: 31.68 -> 33.57Improved
use DUCL curriculum instead of PPL difficulty-only orderLLaMA3-8B CPT · Proof-Pile-2 downstream averageaccuracy: 32.32 -> 33.57Improved
use DUCL instead of PPL orderLLaMA3-8B CPT · MMLU-EMaccuracy: 37.30 -> 40.36Improved
use DUCL instead of PPL orderLLaMA3-8B CPT · GSM8Kaccuracy: 24.87 -> 25.55Improved
use PPL difficulty-only curriculum instead of random orderLLaMA3-8B CPT · validation loss convergencesteps to equal loss: 2600 steps vs 3000 steps for RANDOM; DUCL 2000Hurt
use DUCL instead of random/PPL for faster convergenceLLaMA3-8B CPT · validation losstraining steps: 3000 -> 2000 stepsImproved
use DUCL instead of random orderQwen2.5-7B-Instruct SFT · MATH500accuracy: 80.4 -> 81.6Improved
use DUCL instead of random orderQwen2.5-7B-Instruct SFT · AIME24accuracy: 23.3 -> 26.7Improved
use DUCL instead of PPL orderQwen2.5-7B-Instruct SFT · AIME25accuracy: 16.7 -> 23.3Improved
use DUCL instead of random orderQwen2.5-7B-Instruct SFT · IFEval (forgetting)accuracy: 54.6 -> 57.7Improved
use DUCL instead of random orderQwen2.5-7B-Instruct SFT · average over 7 benchmarksaccuracy: 53.8 -> 55.2Improved
use DUE scoring with strict ordering (alpha=1.0) instead of PPL ordering (ablation)DUCL · CPT early stageVal. loss: 1.506 -> 1.474Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy