A question can have a complicated answer and still be easy to search. Imagine that the answer requires three facts: a person's appointment date, the institution they joined, and the name the institution used at the time. If one well-ranked biography contains all three, an agent may open one page and finish. The reasoning is multi-hop; the retrieval is not.一道题的答案可以很复杂,搜索过程却很简单。假设答案需要三个事实:某人的任职日期、所加入的机构,以及该机构当时的名称。如果排名靠前的一篇传记恰好包含全部信息,智能体只需打开一次页面就能完成。推理是多跳的,检索却不是。
That distinction matters for supervised agent training. The model learns from the trajectory it sees. A task that looks difficult on paper cannot teach successive search decisions if the teacher solved it with a single shortcut.这一区别直接影响智能体的监督训练。模型学到的是轨迹中真实出现的动作。若教师通过一次捷径检索解题,即使题面看起来很难,也教不会模型连续搜索的决策。
Design the discovery path设计证据的发现路径
In ACTS, we treat answer construction and evidence discovery as connected but distinct design problems. One structure records entity facts and their source pages. Another records how a search action can move from one page, through a clue, to a new source. The connection lets us ask a more useful question: what would the agent need to do to discover the next fact?在 ACTS 中,我们把答案构造与证据发现视作相关但不同的设计问题。一层结构记录实体事实及其来源页面,另一层记录搜索动作如何从当前页面,经由线索到达新来源。连接这两层后,便能问出更有用的问题:智能体需要做什么,才能发现下一条事实?
Source provenance is essential here. A fact that merely appears near another fact in a document is not automatically a good bridge. The clue should be grounded in a source the agent can reach, and the next page should support the fact it is used to introduce. Otherwise the generated path is tidy on paper but brittle in a real browser.来源可追溯性在这里至关重要。两个事实碰巧出现在同一文档,不意味着它们天然构成好的桥接线索。线索应落在智能体能够访问的来源上,下一页面也应真正支持它引出的事实。否则纸面上的路径很漂亮,在真实浏览器里却站不住脚。
Train on observed behavior从真实行为中学习
A designed path is still a proposal. The teacher's tool calls reveal whether that path produced the intended behavior. Did it search again? Did it open a new source? Was the final answer supported? These observations are stronger training criteria than counting the number of facts in the answer alone.设计出的路径终究只是一个设想。教师实际调用工具的记录,才能说明路径是否诱发了预期行为:它有没有继续搜索?有没有打开新的来源?最终答案是否有证据支持?这些观察比单纯统计答案中的事实数量更适合作为训练筛选条件。
This also changes how I read a dataset. I look for the action that each example is supposed to teach, then verify that the action appears in the trace and contributes to the outcome. A good search example is a record of a useful discovery process, not just a difficult question with a correct answer attached.这也改变了我审视数据集的方式。我先问每个样本希望教会模型什么动作,再检查这个动作是否真的出现在轨迹中、是否对结果有贡献。好的搜索样本记录的是有用的发现过程,而不只是附着正确答案的难题。