The Art of Interrogation: Consistency Amplifies Factuality in Spatial Reasoning

Researchers propose a self-supervised reinforcement learning framework that improves large language models' spatial reasoning capabilities through consistency verification rather than labeled data. The approach, which uses geometric and semantic consistency checks across image and text transformations, achieves performance comparable to supervised fine-tuning without ground-truth annotations.

This research addresses a critical limitation in current large reasoning models: their poor performance on spatial reasoning tasks despite general capability advances. Rather than treating spatial reasoning as a knowledge gap requiring expensive supervised training, the researchers argue that these capabilities already exist within pre-trained models but lack proper alignment. This distinction is significant because it shifts the problem from data acquisition to inference optimization.

The study's core innovation is the consistency verifier framework, which uses reward functions to evaluate geometric and semantic consistency across transformations. By applying both visual transformations (image flipping) and textual transformations (object order swapping), the model learns to maintain coherent spatial reasoning without labeled examples. The proposed OT-GRPO algorithm optimizes this self-supervised learning through optimal transport-based matching, representing a technical advancement in reinforcement learning methodology.

For the AI development community, this work demonstrates that expensive labeled datasets may not be necessary for improving specialized reasoning capabilities. The label-free approach reduces training costs while achieving competitive accuracy with supervised methods across diverse domains. This finding has implications for model development efficiency and accessibility for researchers with limited annotation resources.

Looking forward, this research opens pathways for improving reasoning in other domains beyond spatial tasks through consistency-based self-supervision. Success here could influence how future large models are trained, potentially reducing reliance on external data sources and supervised fine-tuning workflows. The technique may become foundational for developing more efficient, self-improving AI systems.

• →Self-supervised learning through consistency verification can match supervised fine-tuning performance for spatial reasoning without labeled data.
• →Large language models already contain spatial reasoning capabilities that require alignment rather than new knowledge injection.
• →OT-GRPO algorithm provides a novel reinforcement learning strategy optimized for pairwise consistency verifiers.
• →The approach generalizes across diverse tasks and domains while reducing annotation and training costs.
• →Consistency-based training using geometric and semantic transformations offers a scalable alternative to supervised learning methods.