Research map / Reinforcement learning / RL from feedback
PretrainZero: Reinforcement Active Learning on Pretraining Data
Xingrun Xing, Zhiyuan Fan, Jie Lou, Guoqi Li, Jiajun Zhang, Debing 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 reinforcement active learning (learned mask generation) instead of random span masks in RLPT (ablation) | PretrainZero · Qwen3-4B-Base overall average | overall average: 35.41 -> 39.61 | Improved |
| replace continued next-token pretraining on Wikipedia with PretrainZero | Qwen3-4B-Base · overall average | overall average: 15.89 -> 39.61 | Improved |
| supervised fine-tuning of masked-span QA pairs on Wikipedia instead of RL | Qwen3-4B-Base · overall average | overall average: 32.36 -> 24.27 | Hurt |
| add active learning RLPT (PretrainZero) on Qwen3-8B-Base (ablation) | Qwen3-8B-Base · overall average | overall average: 40.96 -> 42.78 | Improved |
| add active learning RLPT (PretrainZero) on Qwen3-30B-A3B-MoE-Base (ablation) | Qwen3-30B-A3B-MoE-Base · overall average | overall average: 40.38 -> 43.55 | Improved |
| add active learning RLPT (PretrainZero) on SmolLM3-3B-Base (ablation) | SmolLM3-3B-Base · overall average | overall average: 20.25 -> 23.41 | Improved |
| select high-entropy tokens instead of random tokens for next-token RLPT on Wikipedia (ablation) | Random NTP RLPT · MMLU-Pro | accuracy | Hurt |
| use random masked span instead of random next-token prediction in RLPT (ablation) | Random NTP RLPT · MMLU-Pro | accuracy | Hurt |
| add PretrainZero RLPT before general RLVR post-training (2000 RLPT steps) (ablation) | Qwen3-4B-Base · overall average after RLVR | overall average: 37.90 -> 40.46 | Improved |
| increase RLPT steps from 1000 to 2000 before RLVR (ablation) | PretrainZero · overall average after RLVR | overall average: 39.15 -> 40.46 | Improved |
| use Wikipedia instead of MathPile as RLPT data (ablation) | PretrainZero · MMLU-Pro | accuracy | Improved |
| add PretrainZero RLPT instead of Random RLPT before RLVR (ablation) | Qwen3-4B-Base · MMLU-Pro after RLVR | accuracy: 58.62 -> 61.15 | Improved |
Closest papers by meaning
- RLP: Reinforcement as a Pretraining Objective (ICLR 2026 · similarity 0.93)
- On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models (ICML 2026 · similarity 0.92)
- Reasoning Under 1 Billion: Memory-Augmented Reinforcement Learning for Large Language Models (NeurIPS 2026 · similarity 0.89)
- Absolute Zero: Reinforced Self-play Reasoning with Zero Data (NeurIPS 2025 · similarity 0.89)
- Generalization of RLVR Using Causal Reasoning as a Testbed (ICLR 2026 · similarity 0.89)
- Reinforcement Learning for Reasoning in Large Language Models with One Training Example (NeurIPS 2025 · similarity 0.89)
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.
