微软AutoEnvScaling：终端Agent自动设计RL环境
New impressive work from Microsoft and collaborators.
RL环境自动化生成是提升模型训练效率的关键路径，这篇论文给出了完整的终端Agent闭环方案与量化收益，做RLHF或环境生成的同学值得细读。
New impressive work from Microsoft and collaborators.
微软及其合作者带来了一项令人印象深刻的新成果。
And it's one of my favorite recent topics: RL environments.
这也是我近期最感兴趣的主题之一：强化学习环境（RL environments）。
They show that Qwen3.6-35B-A3B goes from 41.1 to 53.3 on Terminal-Bench 2.1 by training on environments it designed itself.
通过在模型自行设计的环境中进行训练，Qwen3.6-35B-A3B 在 Terminal-Bench 2.1 上的得分从 41.1 提升至 53.3。
AutoEnvScaling gives environment design to a proposer agent that works in a terminal, looks at the rollouts of the model being trained, and builds new RL environments from them. The trained model's next rollouts guide the next round.
AutoEnvScaling 将环境设计任务交给一个在终端中工作的提议代理（proposer agent），该代理会观察被训练模型的 rollout 结果，并据此构建新的 RL 环境。训练模型下一轮的 rollout 结果将指导下一轮的设计。
Building environments in a terminal raises the share of valid environments up to 25 times over prompt-only generation, at up to 75% lower cost per valid environment.
在终端中构建环境可使有效环境的比例比仅通过提示词生成的方式提高多达 25 倍，且每个有效环境的成本最高可降低 75%。
With GPT-5.6-sol or Claude Opus 5 as the proposer, Qwen3.5-35B-A3B scores 4.3 and 3.8 points above matched static environment pools. The gain shrinks when the proposer cannot see solver feedback.
当使用 GPT-5.6-sol 或 Claude Opus 5 作为提议代理时，Qwen3.5-35B-A3B 的得分比匹配的静态环境池高出 4.3 分和 3.8 分。当提议代理无法看到求解器反馈时，这一增益会缩小。
The pipeline also produces environments for software engineering, computer use, and GPU kernel work.
该流水线还能生成用于软件工程、计算机操作和 GPU 内核工作的环境。
https://arxiv.org/abs/2610.14163
Chat with Paper: https://academy.dair.ai/papers/autoenvscaling-automating-the-data-flywheel-with-terminal-agents-2610.14163
与论文对话：https://academy.dair.ai/papers/autoenvscaling-automating-the-data-flywheel-with-terminal-agents-2610.14163
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力