Reproduce “CausalMACE: Causality Empowered Multi-Agents in Minecraft Cooperative Tasks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Qi Chai, Zhang Zheng, Junlong Ren, Deheng Ye, Zichuan Lin, Hao Wang
CausalMACE introduces a causality-driven planning framework for multi-agent systems in Minecraft, enabling more efficient and coherent collaboration on complex open-world tasks. The system leverages causal intervention to refine task dependencies, achieving improved completion rates and operational efficiency compared to prior multi-agent and single-agent methods.
Key insights
- Integrating causal inference, specifically by treating game rules as instrumental variables, allows for explicit modeling and refinement of subtask dependencies, ensuring the task graph aligns with environmental constraints.
- The causal intervention mechanism, using Average Treatment Effect (ATE) calculations, effectively identifies and prunes spurious or incorrect edges in the task dependency graph generated by LLMs.
- Combining a globally coherent, causality-empowered task graph with a workload-balanced agent assignment strategy facilitates more robust and efficient multi-agent collaboration in dynamic, open-world settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.