A feature-diversity-guided token compression pipeline for 3D VLMs keeps 94.7% of original QA accuracy at 128 tokens and runs 1.92x faster than uncompressed LLaVA-3D.
Advances in neural information pro- cessing systems31(2018)
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
3DZip: Spatial-Aware Feature Diversity-Guided Token Compression for 3D Question Answering
A feature-diversity-guided token compression pipeline for 3D VLMs keeps 94.7% of original QA accuracy at 128 tokens and runs 1.92x faster than uncompressed LLaVA-3D.