New Reinforcement Learning Framework Released for Long-Horizon Agent Training

The Qwen team has announced the release of QwenGyre, a reinforcement learning (RL) framework specifically designed to facilitate the training of AI agents on extremely long-horizon (XLong) complex tasks. According to reports from Reddit, the framework introduces two primary architectural components intended to optimize the efficiency of agent training: an Elastic Scheduler and a Trajectory Processor.

Technical Architecture and Efficiency Gains

The Elastic Scheduler is designed to reallocate GPU resources between rollout and training phases as demand fluctuates. This mechanism is intended to ensure that live executions remain uninterrupted. Preliminary data suggests this approach provides a speedup of up to 1.85× compared to “Colocate” methods and 1.78× over “Async” methods.

Complementing this is the Trajectory Processor, which manages branching executions—including sub-agents, retries, and compaction—by converting them into prefix-shared trajectory trees. This process aims to provide bounded, deduplicated training samples. The system is also designed to score partial progress even in instances where a process reaches a timeout.

Benchmark Performance and Scaling

In testing scaled to the Qwen3.8-2.4T-A95B model, which utilizes approximately 700,000 tokens per rollout, the framework demonstrated measurable improvements in task completion. Data indicates that QwenGyre increased the NL2RepoBench score from 52.5% to 58.5% within 48 steps. Furthermore, the framework reportedly reduced end-to-end training time by approximately 44% when compared to the Async method.

The technical details of the framework have been documented in a paper available via arXiv (arxiv.org/abs/2609.33848).

While the release of QwenGyre focuses on technical advancements in AI training, broader discussions on social platforms like Reddit continue to touch upon the hardware and financial ecosystems surrounding the industry. For instance, users have engaged in discussions regarding the risks of purchasing modified hardware, such as 4090 48GB GPUs, through platforms like Alibaba. Additionally, individual investor discussions on the platform have highlighted various financial experiences related to Alibaba stock and margin accounts, though these are anecdotal in nature.