Context
Liu and colleagues introduced SEEDS, a preview-guided evidence admission framework for multi-round search agents. The agent selects which retrieved candidates to inspect in full rather than automatically reading a top-ranked prefix. The authors evaluated nine question-answering benchmarks across local and live-web search and reported accuracy improvements with fewer full-document reads.
What changed
An explicit evidence-admission step was proposed between retrieval and detailed source consumption. In the authors' live-web Search-o1 comparison with QwQ-32B, average accuracy rose from 9.28% to 11.78% across BrowseComp, BrowseComp-zh and HLE while fully fetched pages per question fell from 27.81 to 8.30.
Why it matters
An AI agent may retrieve a page yet never admit its full content into the reasoning context. AI discovery and citation measurement should distinguish retrieval exposure, preview inspection, evidence admission and final answer citations.
Visibility OS interpretation
Visibility OS interpretation: SEEDS supports treating evidence admission as a distinct potential bottleneck in agentic source discovery. It does not establish that ChatGPT Search, Google AI Mode, Perplexity, Gemini or Claude use the proposed policy.
What remains uncertain
Will non-prefix evidence admission generalize across independent models, query mixes and production retrieval stacks? How often does preview-based admission change which sources are ultimately cited? How do added preview and decision costs affect end-to-end latency and tokens? Can these benchmark results be independently replicated?