Research map / Reinforcement learning / RL from feedback
DRIVE: Best Data Scheduling Practices for Reinforcement Learning with Verifiable Reward in Competitive Code Generation
Speed Zhu, Chuheng Zhang, Jianwei Cai, Guang Chen, Lulu Wu, Xiaolong Xu, Xuyun Zhang, Saiyong Yang, Wiggin Zhou
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 two-stage RL (entropy expansion then progressive curriculum) after SFT | Qwen2.5-32B SFT model · Codeforces Weekly (33) | pass@1: 0.115 -> 0.182 | Improved |
| add stage 1 entropy expansion before stage 2 (vs stage 2 only, 32k) (ablation) | two-stage RL · LiveCode 08-11 (166) | pass@1: 0.636 -> 0.699 | Improved |
| add stage 2 hard-sample curriculum after stage 1 (vs stage 1 only) (ablation) | two-stage RL · Codeforces Weekly (33) | pass@1: 0.112 -> 0.182 | Improved |
| use 24k instead of 32k max sequence length in stage 1 (ablation) | RL stage 1 · LiveCode 08-11 (166) | pass@1: 0.676 (32k) vs 0.625 (24k) | Mixed |
| stage 2 trained on hard samples only rather than mixed difficulty (ablation) | RL stage 2 · LiveCode 08-11 (166) | pass@1 | Improved |
| use harder but fewer samples (109 internal) in stage 2 (ablation) | RL stage 2 · LiveCode V6 (175) | pass@1: 0.703 -> 0.743 | Improved |
| stage 2 with 24k max length instead of 32k (ablation) | RL only stage 2 · LeetCode Weekly (32) | pass@1: 0.550 -> 0.296 | Hurt |
| add arena learning data filtering (1.27M to 470k prompts) (ablation) | SFT · LeetCode Weekly (32) | pass@1: 0.558 -> 0.553 | No clear effect |
| double weight of hard samples via classifier on arena-filtered data (ablation) | SFT · LeetCode Weekly (32) | pass@1: 0.553 -> 0.578 | Improved |
| double weight of hard samples via classifier on arena-filtered data (ablation) | SFT · LiveCode V5 (167) | pass@1: 0.603 (basic) -> 0.594 (ours) | Mixed |
| increase rollouts per prompt from 8 to 16 and 64 (ablation) | single-sample RL training · single hard sample | accuracy | Improved |
| train on four hard samples instead of the single sample alone (ablation) | RL · single hard sample | accuracy | Improved |
Closest papers by meaning
- Scheduling Your LLM Reinforcement Learning with Reasoning Trees (ICLR 2026 · similarity 0.91)
- Not Only Where, But When: Temporal Scheduling for RLVR (NeurIPS 2026 · similarity 0.90)
- AceReason-Nemotron: Advancing Math and Code Reasoning through Reinforcement Learning (NeurIPS 2025 · similarity 0.90)
- Resource-Efficient Reinforcement for Reasoning Large Language Models via Dynamic One-Shot Policy Refinement (ICML 2026 · similarity 0.89)
- CodeScaler: Scaling Code LLM Training and Test-Time Inference via Reward Models (NeurIPS 2026 · similarity 0.89)
- The Reasoning Boundary Paradox: How Reinforcement Learning Constrains Language Models (NeurIPS 2026 · 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.
