You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
|**[09/28/2025]** 🎉 Ascend NPU support — see [usage guide](https://alibaba.github.io/ROLL/docs/English/UserGuide/ascend/ascend_usage). |
43
+
|**[10/14/2025]** 🎉 Our Paper released, see [Part II: ROLL Flash -- Accelerating RLVR and Agentic Training with Asynchrony](https://arxiv.org/abs/2510.11345), the code will be released soon. |
44
+
|**[09/28/2025]** 🎉 Ascend NPU support — see [usage guide](https://alibaba.github.io/ROLL/docs/English/UserGuide/ascend/ascend_usage). |
44
45
|**[09/25/2025]** 🎉 Our Paper released, see [RollPacker: Mitigating Long-Tail Rollouts for Fast, Synchronous RL Post-Training](https://arxiv.org/abs/2509.21009)|
45
46
|**[09/24/2025]** 🎉 Support [Wan2_2 Reward FL pipeline](examples/wan2.2-14B-reward_fl_ds/reward_fl_config.yaml). Explore the new capabilities! |
46
47
|**[09/23/2025]** 🎉 ROLL aligns with GEM environment definition, providing agentic Tool Use training capabilities, [ToolUse docs](docs_roll/docs/English/UserGuide/agentic/Tool_Use.md). |
@@ -157,6 +158,7 @@ We are continuously working to expand ROLL's capabilities:
157
158
---
158
159
159
160
## 🏆 Notable work based on ROLL
161
+
-[TaoSR-SHE](https://arxiv.org/abs/2510.07972): Stepwise Hybrid Examination Reinforcement Learning Framework for Taobao Search Relevance, with SRPO (hybrid reward model + offline verifier), diversified data filtering, and multi-stage curriculum learning.
160
162
-[EARL](https://arxiv.org/abs/2510.05943): Efficient Agentic RL Systems for LLMs, introducing a dynamic parallelism selector and a layout-aware data dispatcher to boost throughput, reduce memory and data movement bottlenecks, enabling stable large-scale agentic RL without hard context-length limits.
161
163
-[LiveThinking](https://arxiv.org/abs/2510.07685): Real-time reasoning for AI-powered livestreaming by distilling a 670B teacher LLM to a 30B MoE (3B active) via Rejection Sampling Fine-Tuning, then compressing reasoning with GRPO; delivers sub-second latency and ~30x compute reduction, with gains in response correctness (3.3%), helpfulness (21.8%), and GMV in Taobao Live Digital Live Service.
162
164
-[TaoSR-AGRL](https://www.arxiv.org/abs/2510.08048): Adaptive Guided Reinforcement Learning for LLM-based e-commerce relevance, introducing Rule-aware Reward Shaping and Adaptive Guided Replay to improve long-horizon reasoning, rule adherence, and training stability in Taobao Search; deployed in main search handling hundreds of millions of users.
0 commit comments