Research map / Reinforcement learning / RL from feedback
Prompt Curriculum Learning for Efficient LLM Post-Training
Zhaolin Gao, Joongwon Kim, Wen Sun, Thorsten Joachims, Sid Wang, Richard Yuanzhe Pang, Liang Tan
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 PCL value-model prompt selection | GRPO (Qwen3-8B-Base) · MATH500 | accuracy (%): 86.4 -> 88.2 | Improved |
| add PCL value-model prompt selection | GRPO (Llama3.2-3B-it) · MATH500 | accuracy (%): 56.2 -> 57.8 | Improved |
| add PCL value-model prompt selection | GRPO (Qwen3-1.7B-Base) · DeepScaleR average | accuracy (%): 34.0 -> 34.9 | Improved |
| replace dynamic sampling with PCL | DS (Qwen3-4B-Base) · DeepScaleR average | accuracy (%): 45.8 -> 45.7 | No clear effect |
| add PCL value-model prompt selection | GRPO (Qwen3-4B-Base) · MATH time to best checkpoint | hours: 29.2 -> 14.0 | Improved |
| replace SPEED with PCL | SPEED (Qwen3-8B-Base) · DeepScaleR average | accuracy (%): 45.5 -> 52.0 | Improved |
| add SPEED prompt filtering | GRPO (Qwen3-8B-Base) · DeepScaleR average | accuracy (%): 51.4 -> 45.5 | Hurt |
| select prompts with p(x)=0.5 instead of p(x)=0 (ablation) | GRPO (Qwen3-4B-Base) · DeepScaleR m=512 n=8 | average accuracy (%): 42.9 -> 46.1 | Improved |
| select prompts with p(x)=0.5 instead of p(x)=1 (ablation) | GRPO (Qwen3-4B-Base) · DeepScaleR m=512 n=8 | average accuracy (%): 43.7 -> 46.1 | Improved |
| increase generations per prompt from 32 to 128 at p(x)=0.5 (ablation) | GRPO (Qwen3-4B-Base) · DeepScaleR, m x n=4096 | average accuracy (%): 46.0 -> 44.5 | Hurt |
| use value model instead of 3 rollouts per prompt for difficulty | PCL (MATH) · prompt filtering time per step | speedup: 12.1x | Improved |
| use value model instead of 3 rollouts per prompt for difficulty | PCL (DeepScaleR) · prompt filtering time per step | speedup: 16.9x | Improved |
Closest papers by meaning
- CurES: From Gradient Analysis to Efficient Curriculum Learning for Reasoning LLMs (ICLR 2026 · similarity 0.92)
- Reward Is Enough: LLMs Are In-Context Reinforcement Learners (ICLR 2026 · similarity 0.89)
- Reward-Guided Prompt Evolving in Reinforcement Learning for LLMs (ICML 2025 · similarity 0.89)
- AdaCuRL: Adaptive Curriculum Reinforcement Learning with Invalid Sample Mitigation and Historical Revisiting (AAAI 2026 · similarity 0.89)
- Dialogue for Prompting: A Policy-Gradient-Based Discrete Prompt Generation for Few-Shot Learning (AAAI 2024 · similarity 0.88)
- Curriculum Reinforcement Learning from Easy to Hard Tasks Improves LLM Reasoning (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.
