Research map / Vision-language and multimodal / Multimodal LLMs

SpatialStack: Layered Geometry-Language Fusion for 3D VLM Spatial Reasoning

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Spatial reasoning in VLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SpatialStack, which injects multi-level VGGT geometry features into successive early LLM decoder layers (instead of fusing only the final geometry layer with vision tokens), yielding VLM-SpatialStack models on Qwen2.5-VL-3B and Qwen3.5-4B with state-of-the-art open-source spatial reasoning results.

Problem

VLMs struggle with 3D spatial reasoning, and existing geometry-augmented VLMs fuse only the final-layer geometry-encoder features with vision tokens, discarding hierarchical geometric cues needed for both fine-grained perception and high-level reasoning.

Approach

Stack geometry features from multiple VGGT layers (11, 17, 23) and add each, through a layer-specific merger, as residuals to successive LLM decoder layers (0, 1, 2) on the visual-token positions, rather than merging them into the vision pathway.

Key results

• SpatialStack-5B (Qwen3.5) reaches 85.5% average on CV-Bench, the best among listed open-source models, and 69.14 overall in the cross-benchmark ablation vs 68.52 for fine-tuned Qwen3.5.
• Single-layer geometry injection shows a trade-off: layer 11 gives the best low-level average (66.11) while layer 23 gives the best high-level average (66.36).
• Naive multi-layer fusion (64.92 overall) underperforms the best single-layer variant, motivating the layered LLM-side design.

Surprising findings

• Naive multi-layer geometry-vision fusion underperforms the best single-layer injection, suggesting feature interference.
• The base Qwen3.5 model remains strongest on BLINK-Spatial, and geometry-vision fusion variants drop severely on that benchmark.

Limitations

• Gains on BLINK-Spatial are negative relative to the base Qwen3.5 model.
• Layer selection is based on a small set of VGGT layers and preliminary tests; the layer-4 choice was excluded for poor performance.
• Training uses a relatively small ~200k sample mixture with one epoch.

Experimental setup

Datasets: SPAR-234k LLaVA-Hound-64k VLM-3R ScanNet split VSI-590K

Benchmarks: VSI-Bench CV-Bench SPAR-Bench BLINK Video-MME MMBench TempCompass

Metrics: accuracy mean relative accuracy

Baselines: Qwen3.5 (fine-tuned) VG-LLM Spatial-MLLM Cambrian-S SPAR-8B GPT-4o SpatialRGPT SpatialBot VLM-3R

Builds on: VGGT DeepStack VG-LLM Qwen2.5-VL DPT

Closest papers by meaning

• SpatialRGPT: Grounded Spatial Reasoning in Vision-Language Models (NeurIPS 2024 · similarity 0.92)
• SpaceMind: Camera-Guided Modality Fusion for Spatial Reasoning in Vision-Language Models (CVPR 2026 · similarity 0.91)
• G$^2$VLM: Geometry Grounded Vision Language Model with Unified 3D Reconstruction and Spatial Reasoning (CVPR 2026 · similarity 0.91)
• Abstract 3D Perception for Spatial Intelligence in Vision-Language Models (CVPR 2026 · similarity 0.91)
• HiSpatial: Taming Hierarchical 3D Spatial Understanding in Vision-Language Models (CVPR 2026 · similarity 0.91)
• From Flatland to Space: Teaching Vision-Language Models to Perceive and Reason in 3D (NeurIPS 2025 · similarity 0.90)