Research map / Vision-language and multimodal / Multimodal LLMs

SpatialTree: How Spatial Intelligence Branches Out in MLLMs

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Spatial reasoning in VLMs

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.

Introduces SpatialTree, a four-level cognitive-science-inspired hierarchy and capability-centric benchmark (27 sub-abilities) for spatial intelligence in MLLMs, analyzing cross-ability transfer via SFT and hierarchy-aware auto-think RL.

Problem

Spatial abilities in MLLMs are studied with fragmented task-centric benchmarks, leaving unclear how low-level perception, mental mapping, simulation, and agentic competence relate and transfer to one another.

Approach

Organizes spatial intelligence into a four-level capability tree (perception, mental mapping, mental simulation, agentic competence) grounded in cognitive science, then uses correlation analysis, targeted SFT, and a hierarchy-aware RL reward (auto-think that suppresses reasoning for intuitive perception) to probe and improve the hierarchy.

Key results

• Gemini 3 Flash achieves the best overall SpatialTree-Bench score of 57.8; the best open-source model, Qwen3-VL-235B, scores 40.0.
• Single-ability L1 distance SFT yields +36.0% on zero-shot complex in-the-wild distance questions and +27.1% on robotic arm manipulation, while same-level abilities see mixed or negative changes.
• Full RL with auto-think on Qwen2.5-VL-7B raises the weighted average from 27.5 to 30.8 (+3.6), outperforming full RL with thinking (30.1).

Surprising findings

• Naive RL that encourages extensive thinking helps complex reasoning but hurts intuitive perception tasks such as depth and counting.
• Single-ability L1 SFT often causes negative transfer within the same level, while combined multi-ability SFT yields synergy exceeding individual gains.

Limitations

• RL experiments are preliminary and limited to a single 7B model and selected abilities.
• Several benchmarked models are proprietary, making results dependent on API versions.
• Auto-think gains are modest (+3.6 average) and some sub-abilities (e.g., Relation, Localization) decline.

Experimental setup

Datasets: SpatialEngine-generated data SUN-RGBD Hypersim Matterport3D ArkitScenes LLaVA-Video LLaVA-NeXT-Interleave LLaVA-OneVision

Benchmarks: SpatialTree-Bench

Metrics: accuracy MSE angular error success rate weighted average score

Baselines: GPT-4o Gemini 2.5 Pro Gemini 3 Flash Qwen3-VL-235B SpaceR VST SenseNova-base-Qwen3-VL-8B

Builds on: GRPO VST Qwen2.5-VL VSI-Bench MMSI-Bench

Closest papers by meaning

• From Where Things Are to What They Are For: Benchmarking Spatial-Functional Intelligence in Multimodal LLMs (CVPR 2026 · similarity 0.91)
• Spatial-IQ: Deconstructing Spatial Intelligence via Hierarchical Capability Tests (NeurIPS 2026 · similarity 0.91)
• SpatialScore: Towards Comprehensive Evaluation for Spatial Intelligence (CVPR 2026 · similarity 0.89)
• Scaling Spatial Intelligence with Multimodal Foundation Models (CVPR 2026 · similarity 0.89)
• Does Spatial Cognition Emerge in Frontier Models? (ICLR 2025 · similarity 0.88)
• EgoMind: Activating Spatial Cognition through Linguistic Reasoning in MLLMs (CVPR 2026 · similarity 0.88)