Average Cumulative Reward / Score is a key evaluation metric for agent tasks in the fields of large models and reasoning. The cumulative reward or overall score obtained by an agent in an interactive environment measures the quality of the agent's policy and its adaptability to complex dynamic environments. It is commonly used in agent benchmarks such as WebArena, GAIA, and OSWorld. This metric is typically reported as a dimensionless score. Widely used in relevant benchmarks and papers, it effectively reflects the model's performance in this dimension. Higher values are better.