Untitled.mp4
The AIAgents simulation environment and MineCollab benchmark suite are used in this paper to evaluate large language model (LLM) agents in embodied multi-agent environments. MineCollab offers procedurally generated collaborative cooking, crafting, and building activities, while AIAgents provides high-level tool abstractions for open world agent interaction.
We evaluate prominent LLMs including GPT-4o, Claude 4.5 Sonnet, and LLaMA3 variations over job difficulty and agent collaboration burden. Current LLMs have substantial communication and execution delays when expanded to multi-agent coordination, despite promising potential. Finally, we suggest memory augmentation, planning integration, and community-driven framework growth.
