The authors frame this design around a discovery–verification asymmetry: finding a candidate that satisfies multiple constraints can be costly, while checking a candidate can often be decomposed into tractable constraint-wise checks.
The paper also describes an autonomous context-update tool that compresses interaction history while retaining verified evidence and unresolved constraints, without relying on an external model. BAAI reports a dense 4B model and a 122B-A10B Mixture-of-Experts model, and reports performance across BrowseComp, WideSearch, DeepSearchQA, Humanity’s Last Exam, and other reasoning and tool-use benchmarks. These are claims reported in the paper; the supplied source does not establish independent replication or deployment outcomes.