Research map / Large language models / Instruction and fine-tuning
EDCO: Dynamic Curriculum Orchestration for Domain-specific Large Language Model Fine-tuning
Jing-Cheng Pang, Sun Liu, Chang Zhou, Xian Tang, Haichuan Ma, KUN JIANG, Jianlong Wang, Kai Zhang, Sijie Wu, Haoran Cai, Chenwei Wu, Xubin Li and 1 more
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 entropy-based dynamic curriculum instead of random sampling | Qwen3-4B RLFT · Datacom | accuracy: 40.43 -> 46.96 | Improved |
| use EDCO instead of static curricula (RS, AC) in Wireless RLFT | Qwen3-4B RLFT · Wireless | accuracy: 34.35 (RS) -> 38.70 | Improved |
| use static easy-to-hard curricula (Length, AC, PPL) instead of random sampling in SFT | Qwen3-4B SFT · Wireless | accuracy | Hurt |
| use EDCO in SFT instead of PPL curriculum | Qwen3-4B SFT · Wireless and Datacom | accuracy: +2.0% (Wireless), +3.3% (Datacom) over PPL | Improved |
| use EDCO instead of baselines on medical RLFT | Llama3.2-3B RLFT · MedQA | accuracy: 32.9 (RS) -> 36.7 | Improved |
| replace selection signal from inference entropy to perplexity (Dynamic-PPL) (ablation) | EDCO · Datacom (Qwen3-4B) | accuracy: 41.3 -> 47.0 | Improved |
| use heuristic dynamic update without learnable bandit policy (EDCO vs SEC) | curriculum selection · Datacom (Qwen3-4B) | accuracy: 34.78 -> 47.0 | Improved |
| select highest-entropy samples instead of a moderate-entropy window (ablation) | EDCO · Datacom (Qwen3-4B) | accuracy: 44.78 -> 46.96 | Improved |
| use prefix entropy approximation instead of full-sequence entropy (ablation) | EDCO entropy estimator · entropy estimation cost | seconds per sample: 2.24 -> 0.37 | Improved |
| increase prefix length from 20 to 128 tokens (ablation) | EDCO · Datacom (Qwen3-4B) | accuracy: 43.91 -> 46.96 | Improved |
| add quick-answer prompting to prefix entropy (ablation) | EDCO entropy estimator · rank alignment of prefix and full-sequence entropy | rank correlation | Improved |
| increase curriculum size N from 200 to 600 (ablation) | EDCO · final score | final score: 0.801 -> 0.817 | Improved |
Closest papers by meaning
- Unveiling the Secret Recipe: A Guide For Supervised Fine-Tuning Small LLMs (ICLR 2025 · similarity 0.89)
- EvoLM: In Search of Lost Language Model Training Dynamics (NeurIPS 2025 · similarity 0.89)
- TAIA: Large Language Models are Out-of-Distribution Data Learners (NeurIPS 2024 · similarity 0.88)
- An Emulator for Fine-tuning Large Language Models using Small Language Models (ICLR 2024 · similarity 0.88)
- Boosting Multi-Domain Fine-Tuning of Large Language Models through Evolving Interactions between Samples (ICML 2025 · similarity 0.88)
- LensLLM: Unveiling Fine-Tuning Dynamics for LLM Selection (ICML 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.
