Vision-language models are increasingly used as high-level planners for embodied agents, and one emerging approach is to scale test-time compute to boost capability. A new arXiv paper, 'DIRECT: When and Where Should You Allocate Test-Time Compute in Embodied Planners?' (arXiv:2606.12402), examines this strategy.
The abstract begins by noting this trend, then states that scaling test-time compute 'increases…' — the sentence is cut off in the provided source. The paper's title suggests the authors are concerned with the optimal allocation of compute, implying that indiscriminate scaling may not be beneficial.
Given the truncated abstract, the specific downside is not stated in the available text. However, the framing indicates a need to decide when and where to allocate compute rather than simply scaling it uniformly.