Alibaba's Qwen team is drawing fresh attention this week for QwenGyre, a reinforcement learning framework designed to train AI agents on tasks that run for hours at a time. The system was described in a paper titled "QwenGyre: An Elastic Reinforcement Learning Framework for Training xLong-Horizon Agents," authored by thirteen Qwen researchers and posted to arXiv in late September, with detailed coverage following across the AI press on October 12 and 13.
The framework targets a category of workloads the team calls "xLong-horizon" tasks, where a single execution spans hours, involves hundreds of interactions between the model and its environment, and consumes close to a million tokens per rollout. Coding agents that work through an entire repository, or assistants that carry out multi-step operations on a user's behalf, are typical examples. Training models to do this well with online reinforcement learning has been one of the harder infrastructure problems in the field — and it is a natural priority for the team behind Qwen's widely used open-weight agent models. For readers tracking breaking AI news on training infrastructure, QwenGyre is a look at how one of the largest open-model teams has re-engineered its RL pipeline around agentic workloads.
Two Bottlenecks in Long-Horizon RL
The paper identifies two fundamental problems with applying online reinforcement learning to extremely long agent tasks.
The first is GPU idling. Because rollouts vary wildly in execution time — one agent run might finish in minutes while another churns for hours — the training side of the pipeline is constantly waiting on stragglers. The result, according to the authors, is massive idle time across expensive GPU fleets, which are the single largest cost in any RL training run.
The second is trajectory redundancy. Long agent tasks do not proceed in a straight line. They branch: an agent tries one approach, backtracks, tries another, and the branching multiplies as tasks get longer and more complex. Much of this branching history is redundant, and training on all of it wastes compute and destabilizes learning. The authors describe this redundancy as crippling training efficiency at scale.
How QwenGyre Addresses Them
QwenGyre is an end-to-end framework for long-horizon online RL built around two core mechanisms.
The first is elastic GPU reallocation. The framework shifts GPUs between the rollout side (where agents run and generate experience) and the training side (where the model is updated) without interrupting live executions. In conventional setups, the split between rollout and training resources is fixed, so one side idles while the other works. Making the boundary elastic means a rollout-heavy phase can borrow training hardware, and vice versa, as the workload shifts.
The second is a trajectory processor that tames the branching problem. It reconstructs branching histories so that shared prefixes are not treated as unrelated data, scores partial progress on incomplete attempts — important when most long rollouts end in failure — and deduplicates redundant paths to bound training costs.
The economics here are straightforward. In reinforcement learning for large models, compute is the dominant expense, and the rollout–training split determines how much of a GPU fleet is actually earning its keep at any given moment. A framework that keeps GPUs on useful work instead of idling effectively multiplies the throughput of the same hardware. For long-horizon agent training — where the paper describes idling as "massive" — the difference between a static and an elastic split can decide whether a training program is practical at all.
Why It Matters for the Agent Ecosystem
Agent products have moved from demos to deployments, and the training techniques behind them are racing to keep up. A coding agent that completes a real ticket, or an assistant that manages a multi-step workflow, may need hundreds of interactions before it reaches an outcome a human would accept — and reward signals for those outcomes are sparse and hard to define. Online reinforcement learning remains the dominant approach for sharpening agent behavior, which makes the throughput of the RL pipeline a direct constraint on how quickly a lab can improve its agents.
That is why frameworks like QwenGyre matter beyond Alibaba. If long-horizon RL becomes dramatically cheaper per capability point gained, more teams can afford to train agents on the long, messy tasks that resemble real work, rather than on short synthetic episodes that are cheap to roll out but weak as a proxy. The teams that solve the throughput problem will be able to iterate on agentic capability far faster than everyone else.
The Reported Results
The most prominent result in the paper comes from scaling QwenGyre to Qwen's flagship model, Qwen 3.8 2.4T, with rollouts of 700,000 tokens. Training with the framework yielded a 6.0 percent absolute gain on NL2RepoBench — a benchmark for translating natural language requests into repository-level code changes — lifting the model from 52.5 percent to 58.5 percent in just 48 RL steps.
On efficiency, the team reports that QwenGyre delivers speedups of up to 1.85 times over a colocated RL baseline and 1.78 times over an asynchronous setup, across evaluations spanning diverse training domains. Those are substantial numbers in a field where throughput gains are usually measured in single-digit percentages.
Caveats and Availability
As with any lab-published result, the numbers are self-reported by the Qwen team and have not yet been independently reproduced. One practical question for outside researchers is access: while some coverage has described QwenGyre as open source, a clear public code release had not surfaced at the time of writing, so independent verification of the reported speedups is not yet possible. AI Buzz Wire will update as more details become available.
Even so, the paper is a useful signal of where agent training is heading. As agents take on longer and more valuable tasks, the infrastructure required to train them is diverging sharply from the chat-era RL pipeline — and the gap between the two will only grow.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
