腾讯提出ContextPilot:通过细粒度奖励分配优化Agent上下文管理
Interesting paper from Tencent.
Agent上下文管理是提升长程任务稳定性的核心痛点,这篇论文提出的细粒度奖励分配机制提供了可复用的训练思路,值得研究Agent架构的同学关注。
Interesting paper from Tencent.
来自腾讯的一篇有趣论文。
Tencent trains an agent to manage its own working context, and assigns credit at the level of individual context edits.
腾讯训练了一个智能体来管理其自身的工作上下文,并在单个上下文编辑的层级上分配信用(credit)。
Long-horizon tasks force a model to retrieve, integrate and maintain scattered information across many turns, and keeping every interaction history makes the working context grow without bound.
长周期任务迫使模型在多轮交互中检索、整合并维护分散的信息,而保留每一次交互历史会导致工作上下文无限增长。
Recent proactive methods let a model edit its own context with tools, but the toolset stops at search, deletion and summarization.
近期的主动式方法允许模型使用工具编辑其自身上下文,但工具集仅限于搜索、删除和摘要。
ContextPilot adds global planning, long-term memory and adaptive soft compression, so the agent can offload information rather than only discard it.
ContextPilot 增加了全局规划、长期记忆和自适应软压缩能力,使得智能体能够卸载信息,而不仅仅是丢弃它。
The training side is where it gets interesting. Standard RL hands the final trajectory reward to every intermediate edit equally. ContextPilot uses context and entropy variation to find which editing decisions actually mattered, samples branches at those points, and estimates action-level advantages from all branched trajectories passing through that edit.
训练方面是本文有趣之处所在。标准强化学习将最终轨迹奖励平等地分配给每一个中间编辑步骤。ContextPilot 利用上下文和熵的变化来找出哪些编辑决策真正产生了影响,在这些关键点采样分支,并从经过该编辑的所有分支轨迹中估计动作级别的优势值。
On long-context QA and deep search it beats existing baselines across several base models while holding a more compact working context. Code is available.
在长上下文问答和深度搜索任务上,它在保持更紧凑工作上下文的同时,在多个基础模型上超越了现有基线。代码已开源。
Paper: https://arxiv.org/abs/2608.28476
论文:https://arxiv.org/abs/2608.28476
Chat with Paper: https://academy.dair.ai/papers/contextpilot-teaching-agents-for-proactive-context-management-via-fine-grained-r-2608.28476
与论文对话:https://academy.dair.ai/papers/contextpilot-teaching-agents-for-proactive-context-management-via-fine-grained-r-2608.28476
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