Context
An original preprint evaluates ReAct-style iterative retrieval against conventional dense retrieval on ViDoRe v3 and BRIGHT. With the same embedding model, the authors report an average 8.7-point improvement in nDCG@10 but much greater latency and token consumption.
What changed
The study explicitly measures the trade-off between agentic evidence exploration and retrieval cost: 107.4 seconds per query for the tested agentic pipeline versus 0.67 seconds for standard retrieval, plus 764.1K input and 5.8K output tokens per query.
Why it matters
Source discovery in multi-step agents can differ from a single semantic similarity ranking, but additional query fan-out and reasoning impose material latency and cost. Evaluation of AI discovery pipelines should track evidence quality and search budget together.
Visibility OS interpretation
Visibility OS interpretation: Better retrieval from repeated search is not free, and one-shot vs multi-step retrieval comparisons need search-budget controls. The results do not demonstrate how any commercial AI search engine selects citations.
What remains uncertain
Do the reported quality gains persist against query-budget- and token-budget-matched baselines? Can parallel retrieval reduce latency without harming relevance? How well do results transfer to public-web citation systems?