Trending storyDeveloping

Qwen introduces QwenGyre, an elastic RL framework for long-horizon agent training

2 articles1 sourcesince Oct 12Last article 2h ago · Oct 12, 9:11 AM ET

Overview

AISummary of 2 articles

Qwen introduces QwenGyre, an elastic reinforcement learning system for extreme-long-horizon agent training, and says it addresses two challenges of that setting.

The system includes an Elastic Scheduler that reallocates GPUs between rollout and training. Qwen reports speedups of up to 1.85x over Colocate and 1.78x over Async.

Qwen says that, scaled to Qwen3.8-2.4T-A95B with about 700K tokens per rollout, QwenGyre raises NL2RepoBench from 52.5% to 58.5% in 48 steps and cuts end-to-end training time about 44% versus Async. The source for these figures is Qwen's own post, which links to a paper; the post does not specify the two challenges it addresses.

Written by AI from the articles below · updated Oct 12, 9:43 AM ET

Check the sources:

Latest developmentOct 12, 9:11 AM ET

QwenGyre tackles two challenges in long-horizon agentic RL

Article timeline

The articles in this story. Times are ET.

Oct 12

9:11 AM
Qwen@Alibaba_QwenOfficial
QwenGyre tackles two challenges in long-horizon agentic RL
AIQwen says QwenGyre addresses two challenges of long-horizon agentic reinforcement learning. The post does not specify the challenges or results.
9:11 AM
Qwen@Alibaba_QwenOfficial
Qwen releases QwenGyre, an elastic RL framework for long-horizon agent training
AIQwen introduces QwenGyre, an elastic reinforcement learning system for extreme-long-horizon agent training, with an Elastic Scheduler that reallocates GPUs between rollout and training. The system reports speedups of up to 1.85x over Colocate and 1.78x over Async. Scaled to Qwen3.8-2.4T-A95B with about 700K tokens per rollout, QwenGyre lifts NL2RepoBench from 52.5% to 58.5% in 48 steps, cutting end-to-end training time about 44% versus Async.

Heat trend

Not enough continuous observations to show a trend yet.