Answers are grounded in this paper and cite their sources.
Spatial intelligence is fundamental for understanding and interacting with the physical world. However, existing MLLMs are constrained by language-centric reasoning paradigms and lack dedicated mechanisms for explicit spatial relation modeling. To address this gap, we propose S³-Reasoning, a decoupled slow-thinking mechanism that transforms spatial perception into structured symbolic representations to bridge spatial perception and language reasoning via interpretable structural features. Specifically, it represents scenes in a canonical coordinate frame and performs constraint-based reasoning over explicit spatial relations, generating symbolic reasoning traces that enable models to answer spatial queries through logical reasoning rather than heuristic pattern matching. Experimental results demonstrate that S³-Reasoning achieves state-of-the-art performance on five spatial reasoning benchmarks among reasoning-based spatial MLLMs and substantially outperforms the baseline. It also enhances the spatial reasoning abilities of proprietary MLLMs, including Gemini and GPT, through structured prompting.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: