Research map / Reinforcement learning / RL from feedback
DARTS: Distribution-Aware Active Rollout Trajectory Shaping for Accelerating LLM Reinforcement Learning
Yujie Wang, Siwei Chen, Longzan Luo, Xinyi Liu, Xupeng Miao, Fangcheng Fu, Bin CUI
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 token-level streaming (ablation) | VeRL · Ablation, Qwen2.5-14B | speedup: 1.00x -> 1.09x | Improved |
| add distribution-aware (dual-end) sampling with uniform redundancy (ablation) | VeRL + token streaming · Ablation, Qwen2.5-14B | speedup: 1.09x -> 1.40x | Improved |
| add variance-based adaptive redundancy allocation (ablation) | DARTS sampling · Ablation, Qwen2.5-14B | speedup: 1.40x -> 1.63x | Improved |
| shift dual-end ratio L:K from 1:7 to 4:4 (ablation) | DARTS · Qwen2.5-Math-7B | speedup: 1.45x -> 1.25x | Hurt |
| shift dual-end ratio L:K from 1:7 to 4:4 (ablation) | DARTS · Qwen2.5-3B | speedup: 1.29x -> 1.15x | Hurt |
| shift dual-end ratio L:K from 1:7 to 2:6 (ablation) | DARTS · Qwen2.5-3B | speedup: 1.29x -> 1.35x | No clear effect |
| scale cluster from 16 to 64 GPUs (ablation) | DARTS · Qwen2.5-Math-7B | speedup: 1.21x -> 1.55x | Improved |
| replace VeRL with DARTS on BBH | Qwen2.5-32B · BBH zero-shot | accuracy: 78.1 -> 84.7 | Improved |
| replace VeRL with DARTS on BBH | Qwen3-30B-A3B · BBH zero-shot | accuracy: 85.8 -> 84.4 | Hurt |
| replace VeRL with DARTS on BBH | Qwen2.5-3B · BBH zero-shot | accuracy: 34.7 -> 38.7 | Improved |
| apply a larger explicit length penalty instead of distribution shaping | GRPO · downstream math tasks | performance drop: 2%-7% performance drop | Hurt |
| use DARTS instead of Tail Batching | VeRL-based RL · end-to-end throughput | speedup over VeRL: 1.07x-1.62x -> 1.29x-1.77x | Improved |
Closest papers by meaning
- ROLLVERIFY: BRIDGING EFFICIENCY AND ACCURACY IN LONG-TAIL ROLLOUT REINFORCEMENT LEARNING (NeurIPS 2026 · similarity 0.87)
- Lookahead Tree-Based Rollouts for Enhanced Trajectory-Level Exploration in Reinforcement Learning with Verifiable Rewards (ICLR 2026 · similarity 0.86)
- Rollout Pass-Rate Control: Steering Binary-Reward RL Toward Its Most Informative Regime (NeurIPS 2026 · similarity 0.85)
- Parameter Exploration for RLVR via Variational Learning (NeurIPS 2026 · similarity 0.85)
- Planning, Fast and Slow: Online Reinforcement Learning with Action-Free Offline Data via Multiscale Planners (ICML 2024 · similarity 0.84)
- Horizon Reduction Makes RL Scalable (NeurIPS 2025 · similarity 0.84)
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.
