Research map / Vision-language and multimodal / Multimodal LLMs

Scaling Spatial Intelligence with Multimodal Foundation Models

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Instruction-tuned multimodal assistants · 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.

Scales spatial-intelligence QA data to 8M samples (SenseNova-SI-8M) and continues training InternVL3, Qwen3-VL and Bagel to obtain SenseNova-SI models that set open-source SoTA on several spatial benchmarks.

Problem

Multimodal foundation models remain weak at spatial intelligence, partly because spatially grounded training data is scarce, imbalanced and fragmented; it is unclear how data scaling affects spatial capabilities.

Approach

A data-centric approach: build an 8M-sample spatial QA corpus organized by a five-capability taxonomy (Metric Measurement, Spatial Relations, Mental Reconstruction, Perspective-taking, Comprehensive Reasoning), with extra emphasis on under-represented perspective-taking, and continue training existing MLLMs without changing architectures.

Key results

• SenseNova-SI-InternVL3-8B reaches 68.8% on VSI-Bench, 43.3% on MMSI, and 85.7% on MindCube, reported as state of the art among open-source models of comparable size.
• SenseNova-SI-InternVL3-8B outperforms GPT-5 on Perspective-taking and on Comprehensive Reasoning, while GPT-5 remains stronger on Spatial Relations.
• On the Spatial subset of EmbodiedBench, SenseNova-SI-InternVL3-8B raises the official-prompt success rate from 10.4% (InternVL3-8B) to 16.6%, without any finetuning.

Surprising findings

• Spatial chain-of-thought (GPT-5-annotated, MindCube-style CogMap, or extended CogMap) and GRPO reinforcement learning give limited or negative gains over plain QA-style data scaling on VSI-Bench Object Relative Direction.
• The previous open-source MindCube SoTA model scores 50.7 without any image input, nearly equal to its full-input score, indicating language-prior shortcuts.

Limitations

• Scaling gains show saturation as data grows, suggesting data scaling alone will not reach human-level spatial intelligence.
• Text-based chain-of-thought does not reliably improve spatial reasoning.
• Some general-capability benchmarks show small declines relative to base models, and SenseNova-SI-InternVL3-2B lags on several spatial subsets.

Experimental setup

Datasets: VSI-590K MindCube MultiSpa VSR SPEC GQA ScanNet ScanNet++ MessyTable Ego-Exo4D

Benchmarks: VSI-Bench MMSI-Bench MindCube ViewSpatial SITE BLINK 3DSR EmbSpatial MMBench-En EmbodiedBench

Metrics: accuracy MRA CAA success rate

Baselines: GPT-5 Gemini-2.5-Pro Gemini-3-Pro-Preview VST Cambrian-S SpatialLadder Spatial-MLLM MindCube-RawQA-SFT

Builds on: EASI MindCube VLM-3R GRPO

Closest papers by meaning

• SpatialScore: Towards Comprehensive Evaluation for Spatial Intelligence (CVPR 2026 · similarity 0.92)
• MMSI-Bench: A Benchmark for Multi-Image Spatial Intelligence (ICLR 2026 · similarity 0.92)
• Exploring Spatial Intelligence from a Generative Perspective (CVPR 2026 · similarity 0.91)
• Spatial-IQ: Deconstructing Spatial Intelligence via Hierarchical Capability Tests (NeurIPS 2026 · similarity 0.91)
• From Where Things Are to What They Are For: Benchmarking Spatial-Functional Intelligence in Multimodal LLMs (CVPR 2026 · similarity 0.90)
• Unfolding Spatial Cognition: Evaluating Multimodal Models on Visual Simulations (ICLR 2026 · similarity 0.90)