According to an arXiv abstract (2610.24447), recent 3D large language models (3D-LLMs) achieve strong results on scene question answering after extensive post-training. However, the authors argue that existing benchmarks rarely evaluate how these models respond to referential ambiguity or insufficient information.
Given the paper's title — Do 3D Large Language Models Really Use 3D Evidence? — the central concern is whether strong benchmark numbers reflect genuine use of 3D evidence or something else. The abstract is cut off in the source, so the proposed evaluation method and findings are not available from this announcement.
Because this is a single source, there are no differing accounts to compare. The key takeaway is that the paper identifies a gap in current benchmarking practice for 3D-LLMs.