QwenGyre: elastic RL for extreme-long-horizon agent training
- Source
- x.com
- Date
Introducing QwenGyre, bringing elastic RL to extreme-long-horizon (XLong) agent training. ✨ It launches with two core designs: 🥳 - Elastic Scheduler: reallocates GPUs between rollout & training as demand shifts—live executions never interrupted. Up to 1.85× speedup over Colocate, 1.78× over Async. - Trajectory Processor: turns branching executions (compaction, sub-agents, retries) into prefix-shared trajectory trees with bounded, deduplicated training samples. Partial progress after timeout still gets scored. Scaled to Qwen3.8-2.4T-A95B with ~700K tokens per rollout, QwenGyre lifts NL2RepoBench from 52.5% to 58.5% in 48 steps, cutting end-to-end training time by ~44% vs Async. 🔗 Read the paper:
Context
Qwen's QwenGyre addresses an expensive problem in training AI agents: tasks that take hours can leave GPUs waiting instead of helping the model learn. Training involves both running agents on tasks and updating the model from their results. QwenGyre shifts GPUs between those jobs as demand changes. It redirects model requests while keeping the 's workspace and tool execution state in place, allowing available capacity to join training while unfinished tasks continue. The paper explains this scheduling mechanism.
The authors report that QwenGyre completed 48 training steps for Qwen3.8-2.4T-A95B in 75.42 hours, versus 134.55 hours with Async, which reserves separate GPU pools for task execution and training. That is about 44% less time under equal GPU budgets, with comparable training scores. The experiment used NL2RepoBench, a for building software repositories from written specifications. This supports faster training for the tested workload, as reported by the authors. Experimental setup and results.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
- mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
Checking sign-in…
Loading comments…