Sequential-decision research shows search policy and retriever jointly shape evidence exposure

Context

Huo and colleagues introduce a sequential-decision framework for LLM-guided search using select and spawn actions, and compare high-recall retrieval policies across four retrieval pipelines on TREC Deep Learning 2019. Both search policy and underlying retriever materially affected recall in their controlled experiments.

What changed

The authors formalized when a search agent explores an existing ranked list versus generating a new ranking through query reformulation, and evaluated ranking methods and exploration policies under shared recall objectives.

Why it matters

AI agents may expose or miss relevant evidence depending on query-branch decisions and allocation of attention across ranked lists, not just the rank of a webpage for a single query.

Visibility OS interpretation

Visibility OS interpretation: AI discovery should be measured as query spawning, ranked-list exploration, document examination and eventual citation. This is a controlled evaluation framework, not a disclosure of production ChatGPT, Gemini, Google AI Mode or Perplexity retrieval internals.

What remains uncertain

Do select/spawn policies generalize to live heterogeneous web corpora and human-validated relevance? How do these stages affect final citation selection?

Evidence claims

supports

Huo and colleagues (2026) formalized multi-step LLM-guided search as selecting the next document from an existing ranked list or spawning a new ranked list through query reformulation or reranking.

supports

In TREC Deep Learning 2019 experiments, Huo and colleagues reported that retrieval pipeline and exploration policy both materially affected high-recall performance; query reformulation with Move-to-Front achieved the highest recall in their reported policy comparisons, with judgment-cost tradeoffs.