Alibaba's QwenGyre Cuts Long-Horizon Agent Training Time by Up to 1.85x
Alibaba Qwen's QwenGyre elastic RL framework trains hour-long, ~700K-token coding agents up to 1.85x faster, lifting Qwen 3.8 2.4T NL2RepoBench scores from 52.5% to 58.5% in 48 steps — but ships no public code.