SCOPE preprint reports higher agent success with fewer training interactions

An arXiv preprint reports that SCOPE, an AI agent training method, improved success on unseen ALFWorld tasks while using 46.7% fewer interactions than GRPO. With Qwen2.5-3B, it reached 89.6% success on those tasks, a gain of 23.2 percentage points. After supervised training, the agent reviews its own task attempts and generates guidance for specific steps. SCOPE applies that guidance only at selected error steps and rewards efficient completion. With the same model, it achieved 97.7% success on ALFWorld and 85.2% on WebShop, outperforming SEED by 5.9 and 6.3 percentage points, respectively. The reported results come from benchmark experiments.