Mini-BEHAVIOR-Gran：揭示指令粒度对语言引导具身智能体的U型效应
人工智能
2026-04-21 v1
摘要
指令粒度是语言引导具身AI研究中重要却受控不充分的变量。现有基准测试通常将每个任务与单个静态指令配对，难以研究同一任务以不同细节程度描述时智能体行为的变化。我们引入Mini-BEHAVIOR-Gran，这是一个用于受控研究指令粒度的新基准，扩展了Mini-BEHAVIOR，为每个任务提供多个指令变体，范围从高层次目标描述到分步骤指导。通过该基准，我们比较了四种用于跨任务粒度量化的候选指标：标记数、实体数、动作动词数和规划宽度，发现宽度与智能体性能关联最为一致。采用宽度进行训练和评估进一步揭示了指令粒度与性能之间的非单调U型关系，呈现出在细粒度和粗粒度两端都存在性能峰值。进一步分析表明，粗粒度性能回升与浅层 grounding 有关，即智能体学习视觉主导策略。
引用
@article{arxiv.2604.17019,
title = {Mini-BEHAVIOR-Gran: Revealing U-Shaped Effects of Instruction Granularity on Language-Guided Embodied Agents},
author = {Sukai Huang and Chenyuan Zhang and Fucai Ke and Zhixi Cai and Gholamreza Haffari and Lizhen Qu and Hamid Rezatofighi},
journal= {arXiv preprint arXiv:2604.17019},
year = {2026}
}
备注
23 pages, Keywords: Language Grounding, Language Granularity, Instruction Following Agent, Width-based Planning Research Area: Multimodality and Language Grounding to Vision, Robotics and Beyond Research Area Keywords: vision language navigation, multimodality, neurosymbolic approaches