Research map / Reinforcement learning / RL from feedback
FastGRPO: Accelerating Policy Optimization via Concurrency-aware Speculative Decoding and Online Draft Learning
Yizhou Zhang, Ning Lv, Teng Wang, Jisheng Dang
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 online draft learning (ablation) | FastGRPO · GRPO on 3 math datasets (average) | end-to-end speedup: 1.74 -> 2.49 | Improved |
| add concurrency-aware drafting and verification (ablation) | FastGRPO · GRPO on 3 math datasets (average) | end-to-end speedup: 2.21 -> 2.49 | Improved |
| speculative decoding with concurrency-aware settings versus vanilla with early termination (ablation) | GRPO generation · GRPO on 3 math datasets (average) | generation speedup: 1.51 -> 2.78 | Improved |
| fixed small-batch speculative config used at larger batch sizes (ablation) | speculative decoding · batch size scaling | speedup: speedup drops below 1.0x at higher concurrency | Hurt |
| online draft learning does not harm general-purpose tasks (ablation) | draft model · MT-bench and HumanEval | speedup ratio: 1.71 -> 1.72 (MT-bench); 2.01 -> 2.05 (HumanEval) | No clear effect |
| online draft learning on SimpleRL-Abel data (ablation) | draft model · GSM8K (in-domain) | average acceptance length: 3.82 -> 4.55 | Improved |
| skip draft pretraining, train draft only online (ablation) | FastGRPO · SimpleRL-Abel-Level3to5, Qwen2.5-7B-Instruct | average accepted length | No clear effect |
| apply FastGRPO to GRPO variants DAPO and GPG | FastGRPO · 3 math datasets, Qwen2.5-7B-Instruct | end-to-end speedup: 2.34 (DAPO), 2.57 (GPG) | Improved |
| use off-the-shelf speculative decoding (EAGLE-2, HASS, EAGLE-3) in GRPO | GRPO generation · 3 math datasets | end-to-end speedup: average E2E 1.12 to 1.27 | No clear effect |
Closest papers by meaning
- Learning To Draft: Adaptive Speculative Decoding with Reinforcement Learning (ICLR 2026 · similarity 0.91)
- Bridging Draft Policy Misalignment: Group Tree Optimization for Speculative Decoding (ICLR 2026 · similarity 0.89)
- EfficientRollout: System-Aware Self-Speculative Decoding for RL Rollouts (NeurIPS 2026 · similarity 0.89)
- Performance-Driven Policy Optimization for Speculative Decoding with Adaptive Windowing (NeurIPS 2026 · similarity 0.89)
- TIC-GRPO: Provable and Efficient Optimization for Reinforcement Learning from Human Feedback (NeurIPS 2026 · similarity 0.88)
- GVPO: Group Variance Policy Optimization for Large Language Model Post-Training (NeurIPS 2025 · 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.
