Research map / Vision-language and multimodal / Multimodal LLMs

InternSpatial: A Comprehensive Dataset for Spatial Reasoning in Vision-Language Models

ICLR 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 InternSpatial, a 12M-QA open-source spatial reasoning SFT dataset with 19 instruction formats and multi-view rotation estimation, plus InternSpatial-Bench; LoRA-tuned InternVL2.5-8B gains 12.1% on the bench and 10.7% on VSI-Bench.

Problem

Existing spatial reasoning VQA resources for VLMs are limited in scale, scene diversity, instruction format expressiveness, and multi-view supervision.

Approach

A modular, largely template-based data engine that lifts 2D annotations into a canonical view space, expands each QA into up to 19 instruction formats (visual marks and textual references), and adds a rotation angle prediction multi-view task.

Key results

• InternVL-Spatial-8B averages 71.0 on InternSpatial-Bench, +12.1 over InternVL2.5-8B (58.9) and above GPT-4o (61.6).
• InternVL-Spatial-8B averages 52.3 on VSI-Bench, +10.7 over InternVL2.5-8B (41.6).
• General benchmarks stay roughly flat: MathVision +1.8, OCRBench -0.1, TextVQA +0.9, ChartQA -1.6, MMStar +0.2.

Surprising findings

• Rotation estimation remains hard: even the best model reaches only 33.6 on that task, and the spatial-trained model gains just +5.1 there.
• InternVL-Spatial-Raw-8B, trained without format expansion, still improves across all instruction styles, suggesting cross-format transfer.

Limitations

• Template-based generation may underrepresent the richness of natural language in real QA.
• Benchmark counting uses a last-number extraction heuristic and GPT-4o scoring for quiz-style questions, which adds evaluation noise.
• Rotation estimation and multi-view evaluation are limited; the InternSpatial-Bench rotation task has only 409 QAs.

Experimental setup

Datasets: InternSpatial ScanNet MultiScan R2R Objaverse COCO AS-1B Visual Genome 3RScan Cityscapes SpatialRGPT-Bench SpatialBench

Benchmarks: InternSpatial-Bench VSI-Bench MathVista OCRBench TextVQA ChartQA MMStar MathVision

Metrics: accuracy relative error GPT-4o-assigned score

Baselines: InternVL2.5-8B GPT-4o Claude-3.7-Sonnet Llama-4-scout Qwen2.5-VL-72B Gemini-1.5 Pro Gemini-1.5 Flash LLaVA-OneVision-72B LLaVA-NeXT-Video-72B VILA-1.5-40B

Builds on: InternVL2.5 SpatialRGPT SAM2 VSI-Bench

Closest papers by meaning

• Spatial-DISE: A Unified Benchmark for Evaluating Spatial Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.93)
• SpatialRGPT: Grounded Spatial Reasoning in Vision-Language Models (NeurIPS 2024 · similarity 0.93)
• Is A Picture Worth A Thousand Words? Delving Into Spatial Reasoning for Vision Language Models (NeurIPS 2024 · similarity 0.92)
• From Flatland to Space: Teaching Vision-Language Models to Perceive and Reason in 3D (NeurIPS 2025 · similarity 0.92)
• SpatiaLab: Can Vision–Language Models Perform Spatial Reasoning in the Wild? (ICLR 2026 · similarity 0.92)
• MMSI-Bench: A Benchmark for Multi-Image Spatial Intelligence (ICLR 2026 · similarity 0.92)