RAAC study finds deep-research agents repeat searches after evidence discovery saturates

Context

Peer-reviewed EMNLP 2026 research evaluates seven deep-research agent architectures and introduces a Retrieval-Aware Agent Controller (RAAC) to reduce stagnant searches using novelty and information-coverage signals.

What changed

Soudani and colleagues add unsupervised trajectory-level signals for document novelty, successive query diversity and coverage of original query criteria, enabling an agent to continue, redirect or stop its search.

Why it matters

In agentic discovery, repeated search calls do not necessarily expose new documents. Retrieval stopping policies and marginal evidence coverage can affect which sources an agent sees before composing an answer.

Visibility OS interpretation

Visibility OS interpretation: track search novelty, cumulative relevant-document coverage, redundant fan-out and stopping behavior separately from raw search-call count and final-answer accuracy.

What remains uncertain

Do production AI search platforms apply comparable stopping policies? How often does a publisher miss exposure because the agent stops after evidence saturation? How transferable are these results beyond the evaluated benchmarks?

Evidence claims

Emerging Evidenceintroduces

Soudani and colleagues analyze seven deep-research agent architectures and report that many later reasoning and search iterations yield little or no improvement in intermediate retrieval recall or final-answer performance.

Emerging Evidenceintroduces

RAAC uses unsupervised signals for document novelty, successive query diversity, and original-question coverage to choose whether a deep-research agent continues, redirects, or stops searching.

Emerging Evidenceintroduces

On BrowseComp-Plus, Soudani et al. report RAAC reduces search calls by 14 on average across evaluated agents (up to 44 fewer), with answer-accuracy gains up to 10% and 3% on average as stated by the authors.