GRAID: Enhancing Spatial Reasoning of VLMs through High-Fidelity Data Generation

ICLR 2026 Conference Submission16028 Authors

Keywords: Vision Language Models, Spatial Reasoning, Synthetic Data Generation, Human Validation, Fine-tuning, Multimodal Learning

TL;DR: A framework that generates high-quality Spatial Reasoning VQA datasets. Human evals show validity rates in the 90s and far above existing methods

Abstract:

Vision Language Models (VLMs) achieve strong performance on many vision-language tasks but fail at spatial reasoning$\textemdash{}$a prerequisite for many applications. Current training data generation pipelines have significant limitations: single-image 3D reconstruction introduces cascading modeling errors and requires wide answer tolerances, while caption-based methods require hyper-detailed annotations and suffer from generative hallucinations.
We present GRAID, built on the key insight that qualitative spatial relationships can be reliably determined from 2D geometric primitives alone. By operating exclusively on 2D bounding boxes from standard object detectors, GRAID avoids both 3D reconstruction errors and generative hallucinations, resulting in datasets that are of higher quality than existing tools that produce similar datasets as validated by human evaluations. We use our framework to implement 22 question templates spanning spatial relations, counting, ranking, and size comparisons, achieving 91\% human-validated accuracy$\textemdash{}$compared to below 59\% by current methods.
We generate 8.5+ million high-quality VQA pairs using images from the BDD100k, NuImages, and Waymo datasets. Critically, we demonstrate that when trained on GRAID data, models do not simply memorize templates, but rather, learn spatial reasoning concepts that generalize and transfer: models fine-tuned on 6 question types improve on 10+ held-out question types, with accuracy gains of 47.5\% on BDD and 37.9\% on NuImages for Llama 3.2B 11B, and when trained on all questions types, achieve improvements on several existing benchmarks such as BLINK.
We will release the GRAID framework and datasets to accelerate research on VLM spatial reasoning.

Primary Area: datasets and benchmarks

Submission Number: 16028