Robots perform tasks significantly worse when given a medium amount of detail compared to being given almost none.
April 23, 2026
Original Paper
Mini-BEHAVIOR-Gran: Revealing U-Shaped Effects of Instruction Granularity on Language-Guided Embodied Agents
arXiv · 2604.17019
The Takeaway
Instruction granularity has a bizarre U-shaped relationship with robot performance. We usually assume that more detail always helps an agent complete a task. These findings show that a moderate level of instruction causes the model to get confused and fail. It works best with either high-level goals or exhaustive step-by-step scripts. This suggests that AI agents have a specific threshold where they stop thinking and start over-analyzing. Prompt engineers need to avoid the middle ground to keep their agents from stalling.
From the abstract
Instruction granularity is an important yet poorly controlled variable in language-guided embodied AI. Existing benchmarks typically pair each task with a single static instruction, making it difficult to study how agent behavior changes when the same task is described at different levels of detail. We introduce Mini-BEHAVIOR-Gran, a new benchmark for controlled studies of instruction granularity that extends Mini-BEHAVIOR with multiple instruction variants per task, ranging from high-level goal