Today for AI
HOT RADAR
agentHEAT 11.2°

AI CLUSTERED EVENT · 10/12/2026

Alibaba Qwen Releases QwenGyre: Online RL Framework for Long-Horizon Agents Achieving 1.78x Speedup

1 reports archived1 independent sourcesupdated 10/12/2026, 19:20:43
Synthesis & Latest Updates
1 Sources Cross-Validated

Alibaba Qwen introduces QwenGyre, an online reinforcement learning framework designed to optimize training efficiency for long-horizon agent tasks involving millions of tokens. By employing elastic GPU scheduling and trajectory tree processing, it dynamically allocates compute resources without interrupting inference, achieving a 1.78x end-to-end speedup over asynchronous baselines and improving pass rates on benchmarks like NL2RepoBench.

LATEST/Alibaba Qwen introduces QwenGyre, an online reinforcement learning framework designed to optimize training efficiency for long-horizon agent tasks involving millions of tokens. By employing elastic GPU scheduling and trajectory tree processing, it dynamically allocates compute resources without interrupting inference, achieving a 1.78x end-to-end speedup over asynchronous baselines and improving pass rates on benchmarks like NL2RepoBench.

HEAT TRENDHourly heat curve

3 fully observed hours
Now
11.2
Peak
11.820:00
24h change
–

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

TIMELINECoverage timeline

Total 1 reports · Latest first
  1. IT之家 · 智能时代T2·78 pts
    • QwenGyre uses elastic scheduling to immediately shift idle GPU units to training while preserving cache states for running tasks, avoiding redundant computation.
    • It records real context via Token-in Token-out and builds trajectory trees, expanding only selected paths as training samples to maximize data utility.
    • The Partial Scoring mechanism allows timed-out but partially completed tasks to provide useful feedback, reducing wasted compute resources.
    • Benchmarks show a 1.78x end-to-end speedup compared to Async methods under the same GPU budget, validated across multiple tasks on Qwen3.6-122B.