Reason, Then Re-reason: Cross-view Revisiting Improves Spatial Reasoning

Researchers propose ReRe, a training-free framework that improves spatial reasoning in egocentric videos by having multimodal AI models first form a hypothesis, then revise it using synthesized novel viewpoints. The approach demonstrates significant performance gains on spatial reasoning benchmarks without modifying existing model architectures.

The research addresses a fundamental limitation in spatial reasoning tasks: single-turn inference forces AI models to resolve geometric ambiguity using semantic priors rather than verifiable evidence. ReRe introduces a two-phase reasoning methodology where multimodal large language models (MLLMs) initially analyze an egocentric video, then reconsider their conclusions when presented with strategically synthesized alternative viewpoints. This mirrors human reasoning, where observing a scene from multiple angles reduces uncertainty and enables more accurate spatial understanding.

The technical contribution centers on a Geometry-to-Video pipeline that renders complementary novel views from predicted 3D geometry. These synthesized views feature elevated, oblique perspectives with broad scene coverage—angles that provide maximum information gain for validating or revising initial hypotheses. Critically, the framework operates at inference time without requiring model retraining, making it broadly applicable to existing open-source MLLMs.

The performance implications are substantial. Evaluations on VSI-Bench and STI-Bench demonstrate that ReRe enables open-source models to match proprietary state-of-the-art systems, effectively democratizing spatial reasoning capabilities. This has downstream applications across robotics, autonomous systems, and embodied AI where egocentric spatial understanding remains crucial but computationally demanding.

The approach reflects a broader shift toward iterative reasoning in AI systems. Rather than expecting single-pass inference to solve complex spatial problems, the framework embraces revisability—a principle increasingly central to advancing model robustness. Future work may explore dynamic view selection and extension to multi-agent spatial reasoning scenarios.

• →ReRe framework enables AI models to revise spatial reasoning conclusions by observing synthesized novel viewpoints from complementary camera angles.
• →Training-free inference-time approach allows immediate application to existing open-source MLLMs without architectural modifications or retraining.
• →Open-source models enhanced with ReRe achieve performance parity with proprietary state-of-the-art systems on spatial reasoning benchmarks.
• →Geometry-to-Video pipeline strategically generates elevated, oblique views that maximize information gain for hypothesis verification and revision.
• →Iterative reasoning methodology demonstrates that revisitable conclusions improve accuracy in egocentric video understanding tasks.