AWS reports small shopping agent matched its teacher on held-out tasks

AWS reports that a Qwen3.5-4B shopping agent trained on successful runs from GLM-5.3 achieved an average success rate of 0.541 on 324 held-out problems, compared with 0.297 for the unmodified model and 0.540 for the teacher using greedy decoding. The method trains the smaller model to reproduce the teacher's reasoning and tool calls, retaining only runs that solve the shopping task. Removing reasoning from the training targets reduced the student's score to 0.415. The results come from a single run, with expected variation of a few points. General instruction-following also weakened: the student's IFEval prompt-strict score fell from 0.819 to 0.741, and its instruction-strict score fell from 0.872 to 0.813.