Omitting reasoning from training data does not harm AI agents on standard benchmarks

Training smaller student AI agents by imitating larger teacher agents usually has them copy both the teacher's internal reasoning steps (thoughts) and its external actions. This is slow and expensive. This paper investigates whether copying the thoughts is actually necessary. The researchers trained student agents in two ways: one including the teacher's thoughts, one without them. They tested both methods on common agent benchmarks where agents interact with environments. They found that students trained without thoughts performed almost as well as those trained with them, and sometimes better. The main finding is that when the teacher's action already reveals the reasoning behind it, copying the thoughts adds little value. This means training can be faster and cheaper by skipping thought supervision when actions are informative enough. The result challenges the standard practice of fully copying teacher trajectories.