MentalThink: Shaping Thoughts in Mental SVG World
This work provides a novel approach to spatial reasoning in MLLMs by enabling executable mental visualization, addressing the bottleneck of implicit visual reasoning in current models.
MentalThink introduces a visual-symbolic reasoning paradigm where MLLMs generate, render, and interpret SVG code as intermediate visual representations for multi-turn reasoning, achieving 55.1% on VSIBench and 76.0% on MindCube, outperforming prior methods.
We introduce MentalThink, a visual-symbolic reasoning paradigm that equips Multimodal LLMs (MLLMs) with an executable mechanism for "mental" visualization. The core of MentalThink is a think-with-SVG pipeline, where the model learns to generate, render, and interpret scalable vector graphics (SVG) code as an intermediate visual representation for multi-turn reasoning. By creating structured vector sketches, the model can externalize spatial hypotheses, inspect them through deterministic rendering, and reason within a constrained geometric space, effectively mimicking the human process of mental imagery. We instantiate this paradigm through a two-stage training framework, combining Supervised Fine-Tuning (SFT) for SVG syntactic alignment with multi-turn Reinforcement Learning (RL) to encourage iterative inspection, revision, and refinement of intermediate visual hypotheses. Extensive evaluations demonstrate that MentalThink achieves superior performance on spatial understanding and reasoning benchmarks (e.g., 55.1% on VSIBench, 76.0% on MindCube), showing that executable vector graphics provide a verifiable visual workspace for dynamic perspective taking, visual reflection, and compositional scene construction.