MentalThink: Shaping Thoughts in Mental SVG World
Summary
MentalThink enables multimodal large language models to perform visual-symbolic reasoning by generating and interpreting SVG code as an intermediate representation for spatial problem-solving.
View Cached Full Text
Cached at: 07/08/26, 06:48 AM
Paper page - MentalThink: Shaping Thoughts in Mental SVG World
Source: https://huggingface.co/papers/2607.03530 Authors:
,
,
,
,
,
,
,
,
,
,
,
,
Abstract
MentalThink enables multimodal large language models to perform visual-symbolic reasoning by generating and interpreting SVG code as an executable intermediate representation for spatial problem-solving.
We introduce MentalThink, avisual-symbolic reasoningparadigm that equipsMultimodal LLMs(MLLMs) with an executable mechanism for “mental” visualization. The core of MentalThink is athink-with-SVG pipeline, where the model learns to generate, render, and interpretscalable vector graphics(SVG) code as an intermediate visual representation formulti-turn reasoning. By creating structured vector sketches, the model can externalize spatial hypotheses, inspect them through deterministic rendering, and reason within a constrainedgeometric space, effectively mimicking the human process ofmental imagery. We instantiate this paradigm through a two-stage training framework, combiningSupervised Fine-Tuning(SFT) for SVG syntactic alignment with multi-turnReinforcement Learning(RL) to encourage iterative inspection, revision, and refinement of intermediate visual hypotheses. Extensive evaluations demonstrate that MentalThink achieves superior performance onspatial understandingand 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.
View arXiv pageView PDFAdd to collection
Get this paper in your agent:
hf papers read 2607\.03530
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2607.03530 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2607.03530 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2607.03530 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
MentalThink: Shaping Thoughts in Mental SVG World
MentalThink introduces a visual-symbolic reasoning paradigm for multimodal LLMs that uses SVG code as intermediate visual representations for multi-turn reasoning. The two-stage training with SFT and RL achieves strong performance on spatial reasoning benchmarks.
4DThinker: Thinking with 4D Imagery for Dynamic Spatial Understanding
4DThinker is a new framework that enables vision-language models to perform dynamic spatial reasoning using 4D latent mental imagery. The paper introduces scalable data generation and novel fine-tuning methods, including 4D Reinforcement Learning, to improve model performance on complex dynamic tasks.
SVoT: State-aware Visualization-of-Thought for Spatial Reasoning via Reinforcement Learning
The paper proposes SVoT, a reinforcement learning framework that generates interleaved, verifiable intermediate states and visualizations for multi-hop spatial reasoning in MLLMs, achieving significant accuracy gains on new benchmarks involving multi-object interactions and numerical reasoning.
SpatialCLI: Learning to Reason With Spatial Tools, Then Without Them
SpatialCLI presents a framework that trains vision-language models to use specialist spatial tools and then internalize those capabilities, boosting Qwen3-VL-8B-Instruct on the MindCube benchmark from 29.3% to 84.6% with tools.
VisualThink-VLA: Visual Intermediate Reasoning for Effective and Low-Latency Vision-Language-Action Policies
VisualThink-VLA introduces a visual intermediate reasoning framework for vision-language-action policies that preserves spatial precision and dramatically reduces latency compared to text-based reasoning, achieving sub-second inference and state-of-the-art success rates on robot manipulation benchmarks.