Researchers target token compression for 3D VLMs
ORCA and 3DZip aim to reduce 3D CT and scene-token overhead without retraining models.
Why it matters
The work addresses a practical bottleneck for 3D VLMs: compute and memory overhead from large spatial token sets. Better compression could make medical imaging and 3D scene reasoning more tractable for language-model pipelines.
The key points
- 1.ORCA targets 3D CT token compression without model changes.
- 2.3DZip targets 3D question answering with spatially aware compression.
- 3.Both papers focus on reducing 3D VLM token overhead.
Two new papers propose token-compression methods for 3D vision-language workloads, where scans or scenes can produce thousands to tens of thousands of visual tokens. ORCA compresses 3D CT tokens using organ guidance and centroid encoding, and is described as training-free and plug-and-play. 3DZip compresses 3D scene tokens through voxelization, feature-diverse anchor selection and spatially constrained merging, with experiments reporting stronger results than existing compression methods on three 3D question-answering benchmarks.
⚡ Try this today
Review ORCA or 3DZip before building 3D VLM systems that pass large CT or scene-token sets into an LLM.
Sources & original reporting
This brief summarizes and links to reporting from the publishers below.
- arXiv cs.AIORCA: ORgan-Centroid Aggregation for Training-Free 3D CT Visual Token CompressionAug 4, 12:00 PM↗
- arXiv cs.LG3DZip: Spatial-Aware Feature Diversity-Guided Token Compression for 3D Question AnsweringAug 4, 12:00 PM↗
- HF Daily Papers3DZip: Spatial-Aware Feature Diversity-Guided Token Compression for 3D Question AnsweringAug 2, 4:00 AM↗
Enjoyed this brief? Get the next one in your inbox.
More in Research
Prior Audit-Repair Context Shifts LLM Verifier Thresholds Toward Leniency
arXiv:2608.
Improving the matrix multiplication exponent with modern optimization and AlphaEvolve
arXiv:2608.
[Paper] Intern-S2-Mobius: Foundation Model with Decoupled Knowledge and Reasoning
We introduce Mobius-v0, an architecture that comprises a globally shared Memory (FFN) that stores knowledge vectors and multiple Reasoners (Self-Attn) that iter