Research map / Reinforcement learning / RL from feedback
BubbleSpec: Turning Long-Tail Bubbles into Speculative Rollout Drafts for Synchronous Reinforcement Learning
Yuhang Xu, Kaibin Tian, Yang Tian, Zhice Yang, Yifeng Yu, Yan Li, Shengzhong Liu, Fan Wu, Guihai Chen
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 BubbleSpec speculative rollout | Verl rollout, Qwen3-1.7B · RL rollout | rollout time (s): 306.2 -> 167.6 (-45.2%) | Improved |
| add BubbleSpec speculative rollout | Verl rollout, Qwen2.5-VL-7B · RL rollout | rollout time (s): 571.9 -> 400.1 (-30.0%) | Improved |
| add BubbleSpec speculative rollout | Verl rollout, Qwen3-1.7B · RL rollout | average decoding steps: 6741 -> 2913 (-56.8%) | Improved |
| add unified attention for speculative verification (ablation) | BubbleSpec, Qwen3-1.7B · RL rollout | rollout time (s): 294.9s -> 167.6s | Improved |
| use suffix tree instead of n-gram matching (ablation) | BubbleSpec, Qwen3-1.7B · RL rollout | rollout time (s): 305.9 -> 167.6 | Improved |
| use suffix tree instead of n-gram matching (ablation) | BubbleSpec, Qwen3-1.7B · RL rollout | acceptance length: 1.83 -> 2.15 | Improved |
| increase pre-generated samples per prompt from 16 to 32 (ablation) | BubbleSpec, Qwen3-1.7B · RL rollout | rollout time (s): 167.6 -> 184.4 | Hurt |
| decrease pre-generated samples per prompt from 16 to 4 (ablation) | BubbleSpec, Qwen3-1.7B · RL rollout | rollout time (s): 167.6 -> 183.8 | Hurt |
| replace Verl with BubbleSpec in RL training | Qwen2.5-VL-7B · AIME24 | accuracy: 80.72 -> 79.58 | No clear effect |
| replace Verl with BubbleSpec in RL training | Qwen2.5-VL-7B · Geo3K (OOD vision) | accuracy: 48.25 -> 51.41 | Mixed |
| replace RollPacker with BubbleSpec | Qwen2.5-VL-7B · RL rollout | average rollout time (s): 655 -> 397 | Improved |
| exploit intra-GPU bubbles of fastest ranks in addition to inter-GPU bubbles (ablation) | BubbleSpec · RL rollout | rollout stability | Hurt |
Closest papers by meaning
- EfficientRollout: System-Aware Self-Speculative Decoding for RL Rollouts (NeurIPS 2026 · similarity 0.92)
- ROLLVERIFY: BRIDGING EFFICIENCY AND ACCURACY IN LONG-TAIL ROLLOUT REINFORCEMENT LEARNING (NeurIPS 2026 · similarity 0.89)
- DORA: A Scalable Asynchronous Reinforcement Learning System for Language Model Training (NeurIPS 2026 · similarity 0.88)
- AREAL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning (NeurIPS 2025 · similarity 0.88)
- TGRL: Temperature-Grouped Reinforcement Learning for Efficient Exploration in LLMs (NeurIPS 2026 · similarity 0.88)
- Reinforcement Learning Finetunes Small Subnetworks in Large Language Models (NeurIPS 2025 · similarity 0.87)
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.
