A new arXiv preprint, P$^3$-VLM, examines how 3D vision-language models (VLMs) are evaluated for grounding and finds a widespread problem: benchmarks depend heavily on bounding boxes as spatial prompts or representation primitives. According to the abstract, this creates a pervasive "semantic leakage" in both evaluation and prompting protocols.
The authors argue that bounding boxes may inadvertently leak semantic information, undermining the validity of grounding tests. To address this, they propose P$^3$-VLM, a point-based alternative that avoids bounding-box reliance. The abstract is truncated, so the source does not include experimental results or a detailed explanation of the mechanism behind the leakage.
As a preprint, the work has not yet been peer-reviewed, and the full methodology and findings remain unavailable from the provided source. The paper is notable for highlighting a potential flaw in existing 3D VLM benchmarks and for offering a concrete design direction to mitigate it.