Researchers report stronger agent training by adding guidance to practice tasks

Researchers report that their training method improved AI agents' performance on two multi-step task benchmarks. Across three student models, the method produced the best average score in every tested setting, with gains of up to 65% in task-goal completion on AppWorld and up to 61% in resolved rate on SWE-bench Verified. The method gives an agent a short, task-specific instruction when its attempts mostly fail, then trains on another attempt. The authors say a competing technique can teach agents to act as though they have information they never received. These results come from a preprint and its reported benchmark tests.