Deep Research has been using the Orchestrator -> Subagents -> Synthesizer loop since the beginning. It's just strange that they'd put a loop benchmark next to actual model benchmarks.

Maybe it's a tune of the base model that works especially well with the subagent loop?