MineExplorer: Evaluating Open-World Exploration of MLLM Agents in Minecraft

Researchers introduced MineExplorer, a Minecraft benchmark for evaluating multimodal language model agents on open-world exploration.

It filters out overly Minecraft-specific atomic tasks and composes them into implicit multi-hop challenges using multi-agent synthesis for task graphs, scenes, and evaluators.

Strong MLLM agents do well on some single-hop tasks, but performance drops as hidden prerequisites and longer trajectories increase complexity.

The benchmark exposes a gap in long-horizon exploration and dependency handling, which matters for assessing real-world embodied AI skills.