MentalThink: Shaping Thoughts in Mental SVG World

2026-07-03

AI Summary

This work addresses the absence of an executable “mental imagery” mechanism in existing multimodal large language models, which limits their capacity for structured spatial reasoning and dynamic viewpoint switching. The authors propose MentalThink, a novel paradigm that externalizes mental imagery through executable SVGs, leveraging vector graphics as an intermediate representation. By generating, rendering, and parsing SVGs under deterministic rendering and geometric constraints, the framework emulates human visual thinking processes. The approach employs a two-stage training strategy: first aligning SVG syntax via supervised fine-tuning, then iteratively refining visual hypotheses through multi-turn reinforcement learning. Evaluated on benchmarks such as VSIBench (55.1%) and MindCube (76.0%), the method demonstrates significant performance gains, validating the efficacy of a vector-graphics-based workspace for spatial reasoning.

Abstract

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.

Problem

• visual-symbolic reasoning
• mental imagery
• spatial reasoning
• multimodal LLMs
• SVG-based representation

Innovation

• mental visualization
• executable SVG
• visual-symbolic reasoning
• multimodal LLMs
• iterative refinement