Experimental Results of Task-Level CoA Planning

These two scenarios are highly representative and authoritative: firstly, as the core adversarial tasks of a national competition, their design was completed by the organizing committee experts, objectively reflecting the mainstream thinking and complexity requirements of current wargaming research; secondly, their adversarial modes provide experimental scenarios for modeling the behavior and verifying decisions of intelligent agents in adversarial environments.

Specifically, in both scenarios, the blue side was represented by an AI agent pre-designed by experts, while the red side was operated by the contestants. In the experiment, the red side was played by a large language model, which generated task-level action plans to verify the capabilities of the large language model in complex situational reasoning and action planning in a standardized and highly comparable wargaming environment.