CAPruner: Conceptual-Adjacent Scene Graph Pruner for Enhancing 3D Spatial Reasoning of Large Language Models

2 min read

CAPruner cuts down a 3D scene graph before an LLM reasons over it for 3D vision-language tasks such as referring-expression grounding and question answering. Feeding the full graph costs too many tokens, and the usual fix (keep only edges between nearby objects) can drop the one relation to the anchor object that the question is about. CAPruner scores each relation by fuzzy semantic relevance to the query combined with spatial proximity, and keeps the top ones. It avoids relation-level labels by supervising, for each node, the summed score of its incident edges. The README shows it built on the 3DGraphLLM data pipeline, pretrained on ScanRefer, ScanQA and Multi3DRefer, with either ground-truth or Mask3D boxes.

Findings

  • The abstract gives no numbers. It claims the pruner keeps the relations that spatial reasoning needs and “substantially” improves LLM performance on 3D-VL tasks. I have read only the abstract and README.

Limits

  • Indoor ScanNet-style scans, static scenes, offline.
  • Pruning depends on the query, so each new question needs a new pass.
  • Nothing in the abstract on graphs that already contain wrong objects or wrong edges.

Relevance

Deciding which edges a task needs is the same problem a robot team has when it can send only part of a graph over a thin link. Feeds language-grounded-planning and 3d-scene-graphs.