meituan-longcat/MineExplorer
收藏资源简介:
MineExplorer基准测试是一个用于评估多模态大语言模型(MLLM)代理在《我的世界》(Minecraft)开放世界探索能力的数据集。它通过过滤那些解决方案严重依赖《我的世界》特定知识的原子任务,以更好地反映一般开放世界推理能力,并围绕ReAct风格的能力框架组织,将原子任务组合成隐式的多跳任务。数据集使用多智能体合成工作流构建,联合设计任务图、沙盒场景和基于规则的里程碑评估器,以确保实例的可靠性。实验表明,开放世界探索仍然具有挑战性,即使强大模型能处理许多单跳任务,但在需要协调隐藏先决条件的较长轨迹上性能会显著下降。数据集包含813个多跳基准场景和一个100个最困难场景的子集,每个场景记录包括场景ID、任务文本、场景描述、命令、里程碑等字段。
The MineExplorer benchmark is a dataset designed to evaluate the open-world exploration capabilities of multimodal large language model (MLLM) agents in Minecraft. It filters atomic tasks whose solutions heavily rely on Minecraft-specific knowledge to better reflect general open-world reasoning capabilities, and is organized around a ReAct-style capability framework that combines atomic tasks into implicit multi-hop tasks. The dataset is constructed using a multi-agent synthetic workflow, where task graphs, sandbox scenarios, and rule-based milestone evaluators are jointly designed to ensure the reliability of instances. Experiments show that open-world exploration remains challenging: even though strong models can handle many single-hop tasks, their performance drops significantly on longer trajectories that require coordinating hidden prerequisites. The dataset consists of 813 multi-hop benchmark scenarios and a subset of the 100 most difficult scenarios. Each scenario record includes fields such as scenario ID, task text, scenario description, commands, and milestones.




