ContextPilot：通过细粒度强化学习训练Agent主动管理上下文
Long-running agents do not just need a bigger context window. They need to learn…
Agent长程任务的核心痛点是上下文爆炸，这篇论文给出了具体的RL训练方案和量化对比，思路清晰且数据扎实，值得做Agent的同学深入阅读。
Long-running agents do not just need a bigger context window. They need to learn what deserves to stay in context at all.
长期运行的智能体不仅需要更大的上下文窗口，还需要学会判断哪些内容值得始终保留在上下文中。
Long-horizon agents work better when context management is a learned policy
当上下文管理成为一种习得策略时，长周期智能体的表现会更好。
Long-running agents keep adding searches, tool outputs, and reasoning to the prompt. Eventually the model spends more tokens trying to separate useful facts from old clutter.
长期运行的智能体会不断向提示词中添加搜索、工具输出和推理过程。最终，模型会花费更多令牌（token）来区分有用事实与陈旧冗余信息。
ContextPilot gives the agent control over that problem. It can plan, save important information to long-term memory, and summarize, compress, or remove old context.
ContextPilot 赋予智能体解决该问题的能力。它可以规划、将重要信息保存至长期记忆，并对旧上下文进行总结、压缩或移除。
Its RL training then teaches which of those context decisions actually help.
其强化学习（RL）训练随后教会智能体哪些上下文决策真正有效。
With Qwen3-8B, ContextPilot-8B-RL averaged 69.40 across 4 long-context benchmarks, versus 45.93 for the same model using a 128K window without context tools.
使用 Qwen3-8B 时，ContextPilot-8B-RL 在 4 个长上下文基准测试中的平均分为 69.40，而同一模型在不使用上下文工具且采用 128K 窗口时的平均分仅为 45.93。
On BrowseComp, its context stayed around 8K–10K tokens per turn while WebExplorer-8B grew toward 30K.
在 BrowseComp 上，其每轮上下文保持在约 8K–10K 令牌，而 WebExplorer-8B 则增长至接近 30K。
Paper's overall recommendation: stop treating the entire conversation history as memory. Give agents a smaller working context they can actively manage, and train them to keep what matters.
论文的总体建议：停止将整个对话历史视为记忆。给予智能体一个较小的、可主动管理的工作上下文，并训练它们保留关键内容。
– arxiv. org/abs/2608.28476
– arxiv.org/abs/2608.28476
Title: "ContextPilot: Teaching Agents for Proactive Context Management via Fine-grained RL"
标题：《ContextPilot：通过细粒度强化学习教导智能体进行主动上下文管理》
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