Mindcraft Research Paper!

The Mindcraft research paper, “Collaborating Action by Action: A Multi-Agent LLM Framework for Embodied Reasoning,” is co-authored by Emergent Garden, Colby (the other Mindcraft developer), and researchers from UCSD, notably Izzy and Aush.

The Mindcraft team uses a simple speech-bubble mod to visualize LLM-controlled agents’ chat messages directly above their avatars inside the Minecraft world.

The Mindcraft paper defines a suite of automated tasks to evaluate LLM-controlled agents: individual and collaborative crafting tasks, cooking scenarios, and blueprint-based construction challenges.

The Mindcraft paper evaluates several frontier models (at the time) controlling multiple bots, focusing on how well they collaborate to achieve shared goals when inventory, ingredients, or building tasks are split between them.

Researchers and practitioners who want to replicate or extend the Mindcraft experiments must set up the benchmark harness, which involves nontrivial infrastructure steps.