AI Paper Atlas

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

ICLR 2026 · Reinforcement learning › RL from feedback · approach: RL for LLM reasoning · 38 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
add PCL value-model prompt selectionGRPO (Qwen3-8B-Base) · MATH500accuracy (%): 86.4 -> 88.2Improved
add PCL value-model prompt selectionGRPO (Llama3.2-3B-it) · MATH500accuracy (%): 56.2 -> 57.8Improved
add PCL value-model prompt selectionGRPO (Qwen3-1.7B-Base) · DeepScaleR averageaccuracy (%): 34.0 -> 34.9Improved
replace dynamic sampling with PCLDS (Qwen3-4B-Base) · DeepScaleR averageaccuracy (%): 45.8 -> 45.7No clear effect
add PCL value-model prompt selectionGRPO (Qwen3-4B-Base) · MATH time to best checkpointhours: 29.2 -> 14.0Improved
replace SPEED with PCLSPEED (Qwen3-8B-Base) · DeepScaleR averageaccuracy (%): 45.5 -> 52.0Improved
add SPEED prompt filteringGRPO (Qwen3-8B-Base) · DeepScaleR averageaccuracy (%): 51.4 -> 45.5Hurt
select prompts with p(x)=0.5 instead of p(x)=0 (ablation)GRPO (Qwen3-4B-Base) · DeepScaleR m=512 n=8average accuracy (%): 42.9 -> 46.1Improved
select prompts with p(x)=0.5 instead of p(x)=1 (ablation)GRPO (Qwen3-4B-Base) · DeepScaleR m=512 n=8average accuracy (%): 43.7 -> 46.1Improved
increase generations per prompt from 32 to 128 at p(x)=0.5 (ablation)GRPO (Qwen3-4B-Base) · DeepScaleR, m x n=4096average accuracy (%): 46.0 -> 44.5Hurt
use value model instead of 3 rollouts per prompt for difficultyPCL (MATH) · prompt filtering time per stepspeedup: 12.1xImproved
use value model instead of 3 rollouts per prompt for difficultyPCL (DeepScaleR) · prompt filtering time per stepspeedup: 16.9xImproved

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