Indoor referring retrieval favors category names: R@10 rises from 0.626 to 0.911 on ScanRefer, while geometry wins only on paraphrases

Under a fixed token budget for candidate retrieval, semantically matching a query to object category names ranks the target object higher than scoring pairwise geometric relations. The prior default was a query-agnostic scene graph whose edges enter the prompt whether the query describes spatial layout or only names a category; pairwise geometry scoring was the alternative path. On the ScanNet indoor referring benchmark ScanRefer gt_val, MiniLM reaches R@10 of 0.911 and BGE-small 0.913, against 0.626 for the pairwise-geometry scorer QCRR and 0.864 for additive noun matching; on Nr3D, MiniLM reaches 0.941 and QCRR 0.634.

Geometry scoring is not useless: it overtakes category matching when the query does not name the target category verbatim. On the 8,660 ScanRefer queries that name the target category verbatim, Cue-only reaches R@10 of 0.756 and QCRR 0.652; the ranking flips once verbatim matching fails. On the 848 non-exact queries, QCRR reaches R@10 of 0.440 against 0.288 for Cue-only, a gap of 15.2 percentage points; that table does not report MiniLM or BGE-small.

Differences in retrieval ranking carry through to downstream grounding accuracy, and placing the target in the serialized graph matters more than recall alone. In the corrected 3DGraphLLM evaluation, random selection and the query-agnostic nearest-neighbor graph reach downstream Acc@0.25 of 0.443 and 0.410, below the semantic selectors. Under a matched object budget, MiniLM reaches 0.673, the noun baseline 0.647, QCRR 0.557, the full graph 0.670 and the oracle 0.697; the GT-in-budget rates are 0.96, 0.94, 0.82, 1.00 and 1.00.

All main-table experiments use ScanNet-derived indoor referring benchmarks, with objects represented as axis-aligned bounding boxes and default budgets of a fixed number of relations and objects matching the 3DGraphLLM evaluation. The end-to-end portion evaluates one released 3DGraphLLM checkpoint and does not measure other backbones. The authors plan to release the protocol, retriever implementations and per-query outputs so new selectors can be swapped in without rewriting the analysis.

Axis-aligned boxes misrepresent elongated or articulated objects, and view-dependent language is only crudely handled in a fixed world frame. The comparison covers two small category-name encoders and one 3DGraphLLM checkpoint, with other scene-LLM backbones untested. The leave-one-out caption bound uses other ScanRefer annotations of the same instance and is not available at deployment. The token-ratio metric depends on the serialization template, so absolute values are not comparable across templates. On CityRefer outdoor validation all methods sit near chance, so those numbers are omitted from the main tables.