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

Researchers propose CAPruner, a scene graph pruning method that enhances how large language models perform 3D spatial reasoning by preserving task-relevant relations rather than relying solely on spatial proximity. The approach combines fuzzy semantic relevance with spatial proximity to identify critical relations, addressing computational inefficiencies in 3D vision-language tasks.

CAPruner represents a meaningful advancement in optimizing how large language models process 3D spatial information. The core problem it addresses is practical: when LLMs reason about 3D scenes, complete scene graphs become computationally expensive and generate excessive token consumption. Rather than indiscriminately removing relations based on distance metrics, the researchers recognized that task-specific relevance matters more than raw proximity, leading to a hybrid approach that balances semantic meaning with spatial positioning.

The broader context reflects growing integration of LLMs into multimodal tasks beyond text. As companies and researchers push LLMs into vision-language domains, efficiency becomes critical for deployment. CAPruner's innovation lies in its training methodology—using aggregated node scores rather than expensive relation-level annotations makes the approach practical for real-world applications. This reduces annotation overhead while maintaining performance.

For developers building 3D vision-language systems, this work offers tangible improvements in both computational efficiency and spatial reasoning accuracy. The method shows measurable performance gains on established 3D-VL benchmarks, making it relevant for applications in robotics, autonomous systems, and spatial understanding tasks. The open-source release accelerates community adoption and enables further refinement.

Looking forward, the field should monitor how efficiently other pruning strategies can be integrated with semantic reasoning. The techniques pioneered here may influence broader architectural decisions in multimodal models, particularly as 3D understanding becomes more prevalent in AI systems.

• →CAPruner combines fuzzy semantic relevance with spatial proximity to identify task-critical relations in 3D scene graphs.
• →The method preserves task-relevant relations while reducing computational costs and token consumption for LLMs.
• →Training uses aggregated node scores, eliminating expensive relation-level annotations.
• →Experiments demonstrate substantial performance improvements on 3D vision-language tasks.
• →Open-source code enables community adoption and further development.