News · 2026-07-30

AI search agents get better when relevance tells them where to look, not what to read

Researchers at Tencent and the Institute of Information Engineering at the Chinese Academy of Sciences have shown that AI search agents perform substantially better when relevance is used to decide where in a corpus to look first, rather than to produce a ranked list of documents to read. Their system reached 84% accuracy on a hard web-research benchmark against 78% for two strong baselines, while using roughly a sixth fewer tool calls than one and a quarter as many as the other.

Key facts

• 84% accuracy versus 78% for both baselines on BrowseComp-Plus with a 100,000-document corpus, using 23.9 average tool calls against 28.7 and 99.1.
• With a stronger underlying model the method reaches 91%, beating the nearest baseline by 9 points.
• Scaling the corpus to 1 million documents, it holds 79% where the nearest baseline falls to 69%.
• Primary source: A New Role for Relevance: Guiding Corpus Interaction in Agentic Search, arXiv, July 27, 2026; code on GitHub under MIT licence.

Relevance ranking is one of the oldest ideas in information retrieval and it was designed for a specific customer: a human with a screen and limited patience. Its job is to put the best ten things at the top, because a person will not scroll past ten. Everything about how retrieval is built and evaluated follows from that.

An agent is a different customer entirely. It does not stop at ten. It reads something, notices what it did not learn, writes a new query, reads again, and keeps going until it has an answer or runs out of budget. For that customer, a ranked list is an odd deliverable - it answers "what should I show?" when the question is "where should I go next?"

The authors' argument, stated in the paper, is that "relevance should guide corpus interaction, rather than merely select its inputs." Concretely, that means using relevance as an ordering prior on traversal: which documents the agent's search tools scan first, and which of the matches within a document get surfaced to the model first. The agent still searches the corpus directly rather than consuming a pre-built list; relevance just stops it from wandering through the least promising material.

What that fixes is a specific failure the paper identifies in current approaches. Ranked-list retrieval conflates how relevant a document looks with how useful its evidence turns out to be - the two are not the same when you are assembling an answer from several sources. Meanwhile agents that scan a corpus directly treat every location as equally promising, so genuinely useful clues surface late or get cut off by context limits. Both problems get worse as the corpus grows, which is exactly what the scaling results show.

The tool-call numbers are the practical headline. Each tool call is a round trip through the model, which means latency and money, and the runaway baseline in this comparison spent 99.1 calls per query to land 6 points lower in accuracy. Getting a better answer while making a quarter as many calls is the kind of result that changes what a system costs to run, not just how it scores. This is the same economic pressure behind other recent work on stopping agents from rediscovering the same repository every session - once agents are the ones doing the reading, wasted reads are the dominant cost.

The result belongs to a broader rethinking of how retrieval-augmented generation should work when a model, not a person, is the reader. Classic RAG fetches a fixed set of chunks and stuffs them into the prompt, which inherits the ranked-list assumption wholesale. Agentic search replaces that with an interactive loop, and this paper is an argument that the retrieval machinery underneath has not caught up with the change.

A complementary release the same week supplies the components. LightOn published DenseOn and LateOn, a pair of 149-million-parameter retrieval models for multilingual, long-context and code search, releasing the models, the training datasets and the training code - the weights under Apache 2.0 and the tooling under MIT. Small, fully open retrievers are the piece an agentic search stack needs and most open efforts have lacked, since the embedding models at the heart of retrieval have tended to be either closed or under-documented.

The caveat is about generality. These results come from one benchmark family with a maximum of a hundred queries per configuration, on corpora that are large but curated. Whether the ordering prior helps as much on messier material - a real company's document store, with duplicates, near-duplicates and stale copies - is untested. The mechanism is intuitive enough that it probably transfers, but intuitive and demonstrated are different claims, and only the second one is in the paper.

Primary source, verified: read the paper → (arXiv 2607.24223)

Ground Truth. (2026, July 30). AI search agents get better when relevance tells them where to look, not what to read. Ground Truth. https://groundtruth.day/news/search-agents-work-better-when-relevance-tells-them-where-to-look-first.html

@misc{groundtruth:search-agents-work-better-when-relevance-tells-them-where-to-look-first,
  title  = {AI search agents get better when relevance tells them where to look, not what to read},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/search-agents-work-better-when-relevance-tells-them-where-to-look-first.html}
}

Topics: retrieval · agents · search · rag · efficiency · research