AI search agents are learning to grade their own dead ends

The next generation of AI search agents is being taught to do something their predecessors could not: tell which of their own searches are actually doing work. Every modern agent fans out a small forest of queries per question, and most of them do not pull their weight; the answer often turns on one or two turns out of a dozen. The new training signal, introduced in the LAPO paper, is built around that asymmetry.

The LAPO method asks, after each search turn, whether the final answer would change if that step were deleted. It replaces the turn and its retrieved evidence with a fixed [DELETE] placeholder, measures how much the policy's confidence in the correct answer drops, and turns that drop into a reward. No extra reward model, no teacher, no judge model. An agreement check keeps only the signals that point the same direction as the raw attribution, so the agent is not graded on noise.

The result, reported on seven knowledge-intensive QA datasets, is an average exact-match score of 0.326, beating the strongest prior self-rewarding baseline by 0.053 points. The deeper shift is what the mechanism implies: agents are starting to learn the value of their own intermediate steps rather than treating every search as equal. The waste was always in the middle of the chain. Now the agent can be trained to skip it.