The Art of Interrogation: Consistency Amplifies Factuality in Spatial Reasoning

2026-06-10

AI Summary

This work addresses the poor performance of large reasoning models on spatial reasoning tasks, which typically rely on costly annotated data. The authors propose the first unsupervised reinforcement learning framework that aligns the model’s internal reasoning process through geometric and semantic consistency verifiers, eliminating the need for ground-truth labels. Key innovations include a novel consistency verification mechanism, the design of an optimal transport–based OT-GRPO algorithm, and the integration of self-supervised learning with multimodal transformations. Experimental results demonstrate that the method achieves accuracy comparable to supervised approaches under label-free conditions and exhibits strong generalization across multiple tasks and domains.

Abstract

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 spatial data from external vision sources or synthetic engines. In contrast, we argue that for many tasks, spatial reasoning capabilities are already present in pre-trained LRMs but require alignment through logical coherence under geometric 2D and 3D constraints. In this work, we propose a self-supervised reinforcement learning (RL) framework that targets the internal reasoning process without requiring ground-truth annotations. By formalizing the notion of consistency verifiers -- reward functions that check for geometric and semantic consistency under transformations -- we demonstrate that models can improve their spatial reasoning abilities. We use both image transformations, like flipping, and textual transformations, like swapping the order of objects in the question, and propose a new optimal transport-based RL strategy, OT-GRPO, which is a minimal-matching variant of group relative policy optimization tailored to pairwise verifiers. We 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 data domains.

Problem

• spatial reasoning
• Large Reasoning Models
• consistency
• geometric constraints
• factuality

Innovation

• consistency verification
• self-supervised reinforcement learning
• spatial reasoning
• OT-GRPO
• geometric constraints