Multi-agent Framework for Time-Sensitive Complementary Collaboration in Minecraft

The article introduces TickingCollabBench, a Minecraft-based multi-agent benchmark designed for time-sensitive complementary collaboration tasks, emphasizing agent heterogeneity, mandatory collaboration, dynamic environments, and strict real-time constraints. The TickingCollab framework enables the generation of diverse environments and utilizes declarative YAML for task specifications, with an automated benchmark generation pipeline that leverages a language model for task drafting and a verifier for feasibility. This work highlights the challenges LLMs face in dynamic settings, particularly in coordinating under partial observability, which is crucial for practitioners developing robust multi-agent systems.