Multi‑view representations enable pretrained image and video vision‑language models to perceive 3D scenes, yet their long visual token sequences make inference expensive. Redundancy stems from repeated observations of the same region across views and from processing all visual tokens at every LLM layer regardless of relevance.
GeoGR introduces a two‑stage framework that compresses both scene representation and visual computation. At the projector level, GeoSemZip merges cross‑view observations into 3D voxel tokens and combines semantic saliency with spatial coverage to build a compact scene representation. Voxel features not selected are aggregated into contextual anchors to preserve information that would otherwise be discarded.
At the LLM level, an effective visual computation window is identified, revealing that query‑conditioned token relevance stays locally stable but varies over longer layer intervals. Guided by this, GroupRoute employs late entry, early exit, and recoverable group‑wise skipping: low‑relevance tokens bypass full LLM layers while retaining their hidden states, and can re‑join the active group at subsequent anchor layers chosen under a computation budget.
For compression‑aware post‑training, a compact dataset of 67K samples is constructed, retaining only 30% of the original training pool. Under this protocol, GeoGR preserves 95.2% and 96.8% of the dense post‑trained performance on LLaVA‑OneVision‑7B and Video‑3D LLM‑7B respectively, while accelerating LLM prefill by $6.8\times$ and $6.1\times$. These results demonstrate GeoGR’s consistent performance‑efficiency trade‑off across 3D VLMs with or without explicit 3D spatial encoding. Code is available at https://github.com/EIT-NLP/GeoGR
Review: GeoGR’s combination of spatial‑semantic token compression and group routing substantially reduces the computational burden of 3D VLMs without sacrificing language‑vision understanding, offering a practical path toward scalable 3D scene modeling.