An AI research team has found a way to make search systems better at their job just by watching how AI agents already use them.
A new paper introduces Navigation-Informed Embeddings, or NIE, a method that turns an agent's own search behavior into training data. When an AI agent searches for information to answer a question, it generates a trace: which documents it looked at, in what order, and where it stopped. NIE treats the document where the agent stopped as a likely-correct answer and uses the documents it passed over along the way as points of comparison. Two versions were tested: NIE-Stop, which uses only the stopping point, and NIE-Path, which also weighs the full search path. Applied to a BGE dense retriever, the approach lifted a key accuracy measure, Recall@20, from 72.2 to 78.0 on an independent benchmark, with the biggest gains on long, multi-step searches.
The appeal here is the price tag: zero. Most retrieval-system upgrades require fresh human-labeled data or synthetic queries generated by another model, both of which cost time and money. NIE instead recycles logs that agentic systems are already producing as exhaust. For any company running AI agents at scale, that is a free improvement sitting in a database somewhere.
It is a narrow result so far, tested on one encoder and a handful of benchmarks, and it will not replace labeled training data outright. But it fits a pattern worth watching: as more software runs through AI agents, the breadcrumbs they leave behind are becoming a training resource in their own right, no humans required.