Search
On October 12, 2026, Ali QianWen officially introduced QwenGyre, an innovative online reinforcement learning framework specifically crafted to tackle ultra-long-horizon agent tasks. This framework leverages elastic scheduling techniques to partition GPUs into discrete computational units. It then dynamically reallocates idle units to training processes while seamlessly preserving the states of ongoing tasks. Through sophisticated trajectory processing, QwenGyre captures contextual information using a Token-in Token-out mechanism and delivers feedback via Partial Scoring. In NL2RepoBench experiments, the model demonstrated significant improvement, with its pass rate surging from 52.5% to 58.5% after just 48 training iterations. Moreover, it achieved a remarkable 1.78x end-to-end acceleration, all within the constraints of the same GPU budget.
-
C114 Communication Network -
Communication Home
