SpaCE framework proves spatial reasoning limits for multi-frame MLLMs

Multi-modal large language models (MLLMs) have shown impressive spatial understanding through large-scale training, but the theoretical foundations of multi-frame spatial reasoning have been unexplored. A new paper introduces SpaCE, a rigorous framework that establishes four main results: an information-theoretic upper bound on spatial reasoning accuracy based on mutual information between multi-frame observations and targets, a sample complexity bound of order Θ(d_eff·K_max/(ε²·δ)) where d_eff is effective spatial dimension and K_max bounds KL divergence, a PAC-Bayes generalization bound under distribution shift, and a formal characterization of the bias-variance trade-off between explicit 3D representations and implicit reasoning approaches. The crossover conditions identify when each paradigm is provably preferable.

The researchers validate their theoretical predictions on three benchmarks—MultiSPA, CA-VQA, and SpatialRGPT—demonstrating that the bounds are empirically tight. Crucially, they find that frame complementarity (rather than sheer number of frames) is the key driver of multi-frame spatial capacity. This provides the first principled theoretical foundation for understanding when, why, and how multi-frame spatial reasoning succeeds in MLLMs. For practitioners, this means clearer guidance on data requirements and architecture choices for spatial AI tasks, and a rigorous lens to evaluate whether adding more frames or switching between explicit and implicit reasoning will actually improve performance.