Research map / Large language models / Instruction and fine-tuning
Tailoring the Training: Difficulty-Aware Learning Strategy Allocation for Large Language Models
Xiaoling Zhou, Shuaiyu Zhou, Zhemg Lee, Tao Chen, Xirui Li, Peng Chen, Jie Jiang, Wei Ye, Shikun Zhang
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 SFT as an allocatable strategy (ablation) | DALSA · Math + MMLU-Pro overall (6 benchmarks) | accuracy: 48.7 -> 54.3 | Improved |
| add RL as an allocatable strategy (ablation) | DALSA · Math + MMLU-Pro overall (6 benchmarks) | accuracy: 50.6 -> 54.3 | Improved |
| add DISCARD strategy for mastered samples (ablation) | DALSA · Math + MMLU-Pro overall (6 benchmarks) | accuracy: 53.8 -> 54.3 | Improved |
| add anti-curriculum weighting (ablation) | DALSA · Math + MMLU-Pro overall (6 benchmarks) | accuracy: 53.7 -> 54.3 | Improved |
| add ALS loss instead of standard CE (ablation) | DALSA · Math + MMLU-Pro overall (6 benchmarks) | accuracy: 53.8 -> 54.3 | Improved |
| add anti-curriculum weighting to vanilla RL (ablation) | RL (GRPO) · Math + MMLU-Pro overall (6 benchmarks) | accuracy: 46.9 -> 48.2 | Improved |
| add ALS loss to vanilla SFT (ablation) | SFT · Math + MMLU-Pro overall (6 benchmarks) | accuracy: 49.0 -> 50.4 | Improved |
| per-sample SFT/RL/DISCARD allocation instead of ReLIFT hard-sample SFT | ReLIFT · Math + MMLU-Pro overall (6 benchmarks) | accuracy: 52.6 -> 54.3 | Improved |
| DALSA vs standalone RL and SFT | RL / SFT · Math + MMLU-Pro overall (6 benchmarks) | accuracy: +7.4 and +5.3 percentage points | Improved |
| DALSA vs RL on response length | RL (GRPO) · Math + MMLU-Pro overall (6 benchmarks) | response length: 2175 -> 3364 | Mixed |
| DALSA vs ReLIFT on Qwen2.5-Math-1.5B | ReLIFT · Math + MMLU-Pro overall (Qwen2.5-Math-1.5B) | accuracy: 36.5 -> 38.2 | Improved |
| DALSA vs ReLIFT on Qwen3-8B | ReLIFT · Math + MMLU-Pro overall (Qwen3-8B) | accuracy: 49.5 -> 51.0 | Improved |
Closest papers by meaning
- Rethinking the Sampling Criteria in Reinforcement Learning for LLM Reasoning: A Competence-Difficulty Alignment Perspective (AAAI 2026 · similarity 0.90)
- On Predictability of Reinforcement Learning Dynamics for Large Language Models (ICLR 2026 · similarity 0.89)
- Resource-Efficient Reinforcement for Reasoning Large Language Models via Dynamic One-Shot Policy Refinement (ICML 2026 · similarity 0.89)
- D²Evo: Dual Difficulty-Aware Self-Evolution for Data-Efficient Reinforcement Learning (ICML 2026 · similarity 0.89)
- Dynamics-Predictive Sampling for Active RL Finetuning of Large Reasoning Models (ICLR 2026 · similarity 0.88)
- Reinforcement Mid-Training (ICLR 2026 · 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.
