What's this about?
A paper on bringing proprietary-level agentic-search ability to open-source models by "distilling" the behavior of a multi-agent search protocol into an open model, closing the gap between closed and open systems on search tasks.
What exactly is "distillation"?
Distillation is training a smaller or open "student" model to imitate the outputs of a stronger "teacher" system. Here the teacher is a multi-agent search setup (several agents coordinating to search and reason); its behavior is captured and used to teach an open model to do the same, so you get much of the capability without the closed system.
Why is this worth noting?
It's directly in the week's open-vs-closed theme: a concrete recipe for open models to catch up on agentic search, one of the more commercially important agent capabilities.