The Art of Interrogation: Consistency Amplifies Factuality in Spatial Reasoning

Current Large Reasoning Models (LRMs) exhibit remarkable general capabilities but significantly underperform in spatial reasoning tasks. Existing approaches treat this gap as a knowledge deficit, relying on supervised fine-tuning (SFT) to ingest labeled data from external vision sources or synthetic engines. In contrast, we argue that for many tasks, spatial capabilities are already present in pre-trained LRMs but require alignment through principles of internal logical coherence. In this work, we propose a self-supervised reinforcement learning (RL) framework that targets the internal Chain-of-Thought (CoT) process without requiring ground-truth annotations. By formalizing the notion of consistency verifiers—reward functions that check for geometric and semantic consistency under transformations like flipping or swapping the order of objects in the question—and optimizing them via our new OT-GRPO strategy, a minimal-consistency matching variant of group relative policy optimization, we demonstrate that models can self-correct their spatial logic. Our results show that this label-free consistency training approaches the accuracy of models trained with ground-truth supervision and achieves similar generalization across diverse tasks and domains.

Current AI systems can solve many complex problems, but they still struggle with understanding space and geometry in images, such as reasoning about positions, distances, or object relationships. Most existing methods try to fix this by giving the models large amounts of labeled training data created by humans or simulation tools. In this work, we show that these spatial reasoning abilities may already exist inside the models, but are not used consistently. Instead of teaching the models new knowledge, we train them to become more logically coherent in how they reason. We develop a new training method where the AI checks whether its own answers remain consistent when an image or question is modified in simple ways, such as flipping the scene or changing the order of objects in the question. Importantly, this approach does not require human annotations or labeled examples. The model learns from self-consistency alone. Our results show that this self-supervised training method can nearly match the performance of systems trained with expensive labeled data, while also generalizing well to new tasks and environments.