腾讯论文：环境进化提升终端智能体强化学习性能
Banger paper from Tencent on environment evolution.
Agent RL 的核心痛点是环境生成与难度控制，这篇论文提供了不依赖智能体反馈的环境进化新范式，且给出了具体的量化提升数据，值得 Agent 开发者深入研读。
Banger paper from Tencent on environment evolution.
腾讯关于环境演进的精彩论文。
Environment supply is becoming the main limit on agent RL. So this is worth a read.
环境供给正成为代理强化学习的主要瓶颈，因此值得阅读。
(bookmark it)
（收藏它）
Agent RL needs a steady supply of environments hard enough to teach something new. Recent methods build them from the weaknesses an agent shows during its own rollouts.
代理强化学习需要稳定供应足够困难的环境以教授新知识。最近的方法基于代理在自身 rollout 过程中表现出的弱点来构建这些环境。
That approach has two problems. The environments inherit the agent's blind spots, so they generalize poorly. And as the agent improves it exposes fewer weaknesses to build from, so the learning signal weakens.
这种方法存在两个问题：环境继承了代理的盲点，导致泛化能力差；而且随着代理能力提升，其暴露出的弱点减少，可供构建的基础变少，从而导致学习信号减弱。
Environment evolution raises difficulty without watching the agent at all. The paper derives three ways to make an environment harder straight from the multi-turn training objective, then applies them generation by generation on a fixed schedule.
环境演进完全不观察代理即可提升难度。该论文从多轮训练目标中直接推导出三种使环境变难的方法，然后按照固定计划逐代应用。
They test the generator before trusting it. Hy4 preview, Claude Opus 5 and GPT-5.6 Sol all do worse on the evolved environments, which confirms the difficulty.
他们在信任生成器之前会先进行测试。Hy4 preview、Claude Opus 5 和 GPT-5.6 Sol 在演进后的环境中表现更差，这证实了难度的提升。
Plain long-horizon RL on Qwen3.6-27B and Qwen3.6-35B-A3B then adds 14.4 and 18.0 points on Terminal-Bench 2.1.
在 Qwen3.6-27B 和 Qwen3.6-35B-A3B 上进行的纯长视界强化学习随后在 Terminal-Bench 2.1 上分别提升了 14.4 分和 18.0 分。
Paper: https://academy.dair.ai/papers/environment-evolution-for-terminal-agents-2609.04128
论文：https://academy.dair.ai/papers/environment-evolution-for-terminal-agents-2609.04128
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