AI News (2026/10/13): Qwen Launches Online Reinforcement Learning Framework QwenGyre
Executive Summary:
The Qwen team at Alibaba launched the online reinforcement learning framework QwenGyre on October 13, achieving progress in optimizing ultra-long-term agent tasks. The framework reduces GPU idle time through an elastic scheduling mechanism and employs trajectory processing technology to control forking training costs, resulting in significant performance improvements in code generation benchmarks.
News Details
QwenGyre was developed to address training bottlenecks in agents handling complex multi-step tasks. By dynamically adjusting computational resource allocation strategies, the framework achieves a 58.5% pass rate on NL2RepoBench within just 48 training steps for a single 700,000-token rollout task using the Qwen3.8-2.4T model, compared to an initial 52.5% pass rate.
Key Features
[Elastic Scheduling Mechanism]
Implements dynamic resource allocation strategies to increase GPU utilization by 36% during approximately 700,000-token rollout tasks, effectively reducing idle time proportions[Trajectory Processing Technology]
Controls forking training costs through branch prediction algorithms, decreasing redundant path computations by 42% in code generation scenarios and resolving resource waste issues in multi-decision-point tasks[Cross-Scale Validation Capability]
Completes cross-task validation testing on a 122B-parameter model, achieving a 1.78× speedup ratio compared to the Async framework under identical GPU budgets while saving 44% of training time
AI-ALL In-Depth Analysis
QwenGyre introduces a new paradigm for training large model agents. Its elastic scheduling design breaks through the rigid constraints of traditional asynchronous frameworks in resource allocation, significantly reducing hardware costs while maintaining inference quality – this "on-demand allocation" model could potentially reshape current reinforcement learning infrastructure systems based on fixed resource allocation.
The trajectory processing module's precise control over forking paths demonstrates industry-leading value. Mainstream Agentic Workflow solutions currently face cost escalation issues caused by decision tree explosions, but this framework reduces computational complexity from O(n²) to O(n log n) through probabilistic pruning strategies, providing reusable technical pathways for multi-step task optimization.
Notably, its effectiveness in cross-parameter validation is significant. Performance on the 122B model indicates strong scalability characteristics, with this parameter-agnostic optimization capability potentially accelerating practical applications of larger models – particularly creating differentiated advantages in high-token-consumption scenarios like code generation.