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

Overview

Large language models can use 3D scene graphs for spatial reasoning, but full graphs are expensive and noisy. Proximity-only pruning is efficient, yet it can remove relations that are essential for a particular 3D vision-language task.

CAPruner addresses this by selecting task-specific scene-graph relations with both semantic relevance and spatial proximity. The goal is to keep a compact relation set that still preserves the key spatial evidence needed for grounding and question answering.

Detailed CAPruner scoring and pruning pipeline — CAPruner scores candidate relations with fuzzy semantic matching and spatial encoding, then prunes the scene graph into a task-specific compact graph.

Key Ideas

Fuzzy matching links object categories with the task context.
Semantic relevance and pairwise distance are combined to score relations.
Edge scores are aggregated into node-level supervision, avoiding costly relation-level annotations.
The inference branch produces a compact task-specific scene graph for downstream LLM reasoning.

Results

CAPruner improves LLM performance on 3D vision-language tasks while reducing unnecessary graph reasoning overhead. The method is designed to keep task-critical relations instead of relying only on nearest-neighbor proximity.

Resources

BibTeX

@misc{zhou2026capruner,
  title={CAPruner: Conceptual-Adjacent Scene Graph Pruner for Enhancing 3D Spatial Reasoning of Large Language Models},
  author={Shengli Zhou and Xiangchen Wang and Guanhua Chen and Feng Zheng},
  year={2026},
  url={https://www.researchgate.net/publication/404391563_CAPruner_Conceptual-Adjacent_Scene_Graph_Pruner_for_Enhancing_3D_Spatial_Reasoning_of_Large_Language_Models}
}