AI Agent 信息缺失时的主动推理策略与成本优化
AI agents have an expensive problem: when a request is missing information, they…
AI agents have an expensive problem: when a request is missing information, they either guess too early or keep asking, retrieving, and calling tools without knowing whether more context is worth the cost.
AI 智能体面临一个高昂的问题：当请求缺少信息时，它们要么过早猜测，要么不断询问、检索和调用工具，却不清楚获取更多上下文是否值得付出相应成本。
This paper uses active inference to make that choice explicit: every clarification, retrieval, tool call, or prompt test should earn its tokens, latency, or user effort by reducing uncertainty that matters to the final answer.
本文利用主动推断（active inference）使这一选择显式化：每一次澄清、检索、工具调用或提示测试，都应通过降低对最终答案至关重要的不确定性，来证明其消耗的 token、延迟或用户努力是合理的。
In controlled tests, frontier models steadily narrowed down hidden answers, yet still used more questions than an optimal planner.
在受控测试中，前沿模型稳步缩小了隐藏答案的范围，但仍比最优规划器使用了更多的提问次数。
On a generation task, targeted clarification raised verifier compliance from 0.0417 to 0.375, while average token use rose from about 112 to 219.
在一个生成任务中，有针对性的澄清将验证器的合规率从 0.0417 提升至 0.375，而平均 token 使用量则从约 112 上升至 219。
The practical recommendation is to give agents an explicit context-acquisition layer instead of treating context gathering as an automatic behavior.
实际建议是为智能体提供一个显式的上下文获取层，而不是将上下文收集视为一种自动行为。
Before each extra step, decide whether to ask, retrieve, inspect, or act.
在每一步额外操作之前，决定是询问、检索、检查还是执行动作。
Get only the missing context that changes the decision, then stop.
仅获取能改变决策的缺失上下文，然后停止。
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