Context
Zhuang and colleagues introduce RISE, a BM25-bounded persistent document workspace that agents can inspect with shell tools and section metadata. In a 100-query BrowseComp-Plus experiment on 100,000 documents, RISE with gpt-5.4-mini reached 78% judged accuracy at $0.28 per query versus 78% and $1.10 for a direct-corpus-interaction baseline.
What changed
The proposed retrieval interface supplies an inspectable candidate-document workspace rather than only a top-k snippet list or unbounded full-corpus shell access. A separate 1-million-document experiment evaluates the boundary-only RISE-BM25 variant, not full RISE.
Why it matters
For AI discovery research, the experiment distinguishes document candidate inclusion, access to evidence inside a retrieved document, and final answer generation. It demonstrates that evidence-delivery interfaces can influence agent effectiveness and cost in controlled tasks.
Visibility OS interpretation
Visibility OS interpretation: measuring AI discoverability should not conflate retrieval ranking with actual evidence exposure and citation. The paper offers an architectural research result, not evidence that any commercial AI search system uses RISE or rewards specific publisher formatting.
What remains uncertain
Can the findings replicate on live public-web search and real publisher citation outcomes? How does full RISE scale beyond 100,000 documents? How sensitive are results to the fixed 100-query sample, model choices, evaluation judge, cost assumptions and differing baseline budgets?