QwenGyre: Elastic RL framework trains xlong-horizon agents with up to 1.85x speedup
aigclink · x · 2026-10-12
A new arXiv paper introduces QwenGyre, an end-to-end online RL framework for training LLM agents on extremely long-horizon tasks, where a single rollout can span hours, hundreds of interactions, and nearly 1M tokens.
Key ideas:
- Elastically reallocates GPUs between rollout and training without interrupting live executions
- A trajectory processor reconstructs branching histories, scores partial progress, and deduplicates redundant paths to bound training costs
Results: scaled to Qwen 3.8 2.4T with 700K tokens per rollout, QwenGyre delivers a 6.0% absolute gain on NL2RepoBench (52.5%→58.5%) in 48 steps, with up to 1.85x and 1.78x speedups over Colocate and Async baselines.
More from coding & agent
- TestSuiteBench: coding agents kill only ~55% of bugs with generated test suites — lbeurerkellner · 2026-10-13
- Running coding agents in YOLO mode: a practical 5-layer guardrail setup — Arindam_1729 · 2026-10-13
- Cloudflare Agents SDK adds web_search (Exa/Linkup/ceramic) and Markdown web_fetch tools — irvinebroque · 2026-10-13
- 45 minutes with AI: failure cases, BQ datasets, and a generative UI — rseroter · 2026-10-13
- DeepSeek desktop agent app reviewed: model promising, harness trails Cursor in 5 ways — casper_hansen_ · 2026-10-13
- Hone, founded by ex-Cognition and Mercor staff, builds business-running agents — nmasc_ · 2026-10-13