arXiv 2608.02751, cs.IR

Search. Inspect. Fetch.

A Boolean search–inspect–fetch method for deep-search agents: the agent skims before it reads. Sieve is one setting of the SkimSearchAgent library.

Shuai Wang1 Haodong Chen1 Yu Yin1 Shengyao Zhuang Bevan Koopman2,1 Guido Zuccon1
1ielab, The University of Queensland 2Australian e-Health Research Centre, CSIRO
Read the paper GitHub

SkimSearchAgent library

SieveBoolean search · fetch sections
$0.00000 · 0 tok · 0 steps
warming up…
Inspect
Fetch §
Answer
this board is driven by a real recorded agent run · not a video
Headline results

Higher accuracy, with a third to half fewer tokens

Against the BM25 Search-Visit baseline under identical budgets (k=5, 12,000-token reads, 100 steps). Skimming structure before reading is not an accuracy-for-cost trade: on all three collections it wins on both axes.

−32.4%
tokens per episode on BrowseComp-Plus, at +2.5 judge points
−50.6%
tokens on MuSiQue, at +3.1 EM. The saving grows where the baseline reads most
3/3
collections where Sieve improves accuracy and cuts cost
BM25 Search-Visit baselineSieve

Accuracy

HotpotQA (7,343)
baseline43.7
Sieve45.3 EM
MuSiQue (2,409)
baseline26.1
Sieve29.2 EM
BrowseComp-Plus (830)
baseline34.7
Sieve37.2 judge

Tokens / episode

HotpotQA (7,343)
baseline19.9k
Sieve13.9k (−30.4%)
MuSiQue (2,409)
baseline43.0k
Sieve21.3k (−50.6%)
BrowseComp-Plus (830)
baseline68.1k
Sieve46.0k (−32.4%)

These are the paper's numbers, in the paper's setting. The library's leaderboard uses a different fixed setting, so its BM25 row is not comparable with the baseline here.

The effect transfers across agent backbones (Qwen-AgentWorld, OpenResearcher), with the largest savings where the baseline burns the most context. Full tables, ablations and statistics are in the paper.

The method

Skim the structure, read only what matters

A deep-research agent that reads whole documents burns its context window on text it never needed. Sieve composes four separable stages on top of one shared agent loop, so the agent narrows before it spends.

  1. 01

    Filter

    The agent writes a BQL Boolean query: field-scoped terms (title, section, date…) with AND/OR/NOT, phrases and prefixes that carve the corpus down to candidates.

  2. 02

    Rank

    Candidates are ranked by relevance, with graceful soft fallback when a strict filter over-narrows to zero hits.

  3. 03

    Inspect

    Each result is a structure-rich card: title, section headings, matched fields, a query-biased snippet. Enough to judge a document without reading it.

  4. 04

    Fetch

    The agent reads only the named section it chose from the card, a slice rather than the whole document, and answers from grounded evidence.

Sieve architecture: filter, rank, inspect, fetch stages over one agent loop. Sieve architecture, dark version: filter, rank, inspect, fetch stages over one agent loop.
The four stages compose over one shared agent loop; each is swappable on its own.
The team

Built at ielab

Portraits of the six authors: Shuai Wang, Haodong Chen, Yu Yin, Shengyao Zhuang, Bevan Koopman and Guido Zuccon. Portraits of the six authors, night version: Shuai Wang, Haodong Chen, Yu Yin, Shengyao Zhuang, Bevan Koopman and Guido Zuccon.
Citation

If you use Sieve or the library

bibtex
@misc{wang2026sieve,
  title         = {Search, Inspect, Fetch: Exploiting Boolean Retrieval
                   for Deep-Research Agents},
  author        = {Wang, Shuai and Chen, Haodong and Yin, Yu and
                   Zhuang, Shengyao and Koopman, Bevan and Zuccon, Guido},
  year          = {2026},
  eprint        = {2608.02751},
  archivePrefix = {arXiv},
  primaryClass  = {cs.IR},
  url           = {https://arxiv.org/abs/2608.02751}
}