ResearchBenchmarksApp-based Agentic Task on AppWorld unseen tasks (test)Follow66.6Pass@1SGE47.04852.12457.262.276Mar 2, 2026Evaluation ResultsMethodMethodLinksPass@1SGERL Training=Strategy-G...RL Training=Strategy-Guided Exploration (SGE), Aggregation=Average ± standard error across 3 random seeds2026.0366.6GRPORL Training=GRPO, Aggr...RL Training=GRPO, Aggregation=Average ± standard error across 3 random seeds2026.0349.3Zero-ShotEvaluation Protocol=Di...Evaluation Protocol=Direct base-LLM evaluation2026.0347.8