MentalThink Equips MLLMs with Visual-Symbolic Reasoning via SVG.

Multimodal Large Language Models (MLLMs) are advancing rapidly, but often lack a robust mechanism for spatial and visual reasoning that mirrors human cognitive processes. Humans frequently use mental imagery to test hypotheses, reflect on visual information, and construct complex scenes. This research aims to equip MLLMs with a similar capability. The paper introduces MentalThink, a novel visual-symbolic reasoning paradigm centered around an executable "think-with-SVG" pipeline. In this framework, MLLMs learn to generate Scalable Vector Graphics (SVG) code, render these graphics, and then interpret them. This SVG code serves as an intermediate visual representation, allowing the model to externalize spatial hypotheses and reason within a structured geometric space over multiple turns. MentalThink employs a two-stage training process: Supervised Fine-Tuning (SFT) for aligning with SVG syntax, followed by Reinforcement Learning (RL) to encourage iterative inspection, revision, and refinement of these visual hypotheses. Extensive evaluations on spatial understanding and reasoning benchmarks demonstrate significant performance improvements, suggesting that executable vector graphics provide a verifiable and dynamic visual workspace for MLLMs.