News from the IT community on October 12th: Alibaba Qianwen officially released QwenGyre today, which is an online reinforcement learning framework designed for long-term intelligent agent tasks.
According to official introductions, when agents undertake complex tasks, a single execution often lasts for several hours. Hundreds of rounds of interdependent interactions occur between the model and the environment, processing millions of token. Through elastic scheduling and trajectory management, such ultra-long-term tasks become an effective source of training, while also leading to improvements in capability and efficiency.
QwenGyre is capable of dividing GPU into several computing units. With each batch of tasks completed, the freed-up units are then allocated to training; tasks that are still running are not disturbed, and their progress and status are preserved. The cached data previously calculated can also be transferred in its entirety, eliminating the need to start over. The newly freed units can be directly incorporated into the current round of training, and ultimately, the core units will aggregate the results to complete a parameter update.
IT noticed that QwenGyre can use Token-in and Token-out to record the true context at the time each output is generated, organizing it into a trajectory tree based on the shared prefix. Only the selected paths are expanded into training samples. At the same time, tasks that time out but are partially completed can also provide useful feedback through Partial Scoring, avoiding the waste of intermediate processes.
In terms of experimental results, the official stated that in the NL2RepoBench online reinforcement learning experiment, each training task took approximately 700,000 token. After 48 rounds of training, the model's pass rate increased from 52.5% to 58.5%. With the same GPU budget, QwenGyre achieved a 1.78-fold end-to-end acceleration compared to Async, and its training scores were close to those of Async and Colocate. The official also mentioned that in the cross-task validation conducted on Qwen3.6-122B, all three types of tasks, namely NL2RepoBench, DeepSWE, and TerminalBench, saw end-to-end speed improvements.
