AI/ ai · multimodal · spatial-reasoning · research

MentalThink Hits 55.1% VSIBench by Drawing Its Own Scratch Work

A new arXiv paper shows multimodal LLMs reason better in 3D space when they sketch SVG diagrams mid-thought, scoring 55.1% on VSIBench and 76.0% on MindCube.

Researchers say they can improve a multimodal AI's spatial reasoning by making it doodle.

A paper posted to arXiv introduces MentalThink, a training framework that teaches multimodal large language models to generate scalable vector graphics (SVG) — essentially simple geometric sketches — as a scratchpad during multi-step reasoning tasks. The model writes SVG code, renders it into an image, inspects that image, then revises its hypothesis. Training runs in two stages: supervised fine-tuning to get SVG syntax right, followed by reinforcement learning that rewards the model for iteratively correcting its own visual guesses. The result is something the researchers liken to human mental imagery — externalizing a spatial hypothesis before committing to an answer.

The benchmark numbers are specific enough to take seriously. MentalThink scores 55.1% on VSIBench and 76.0% on MindCube, both spatial understanding tests where current models routinely struggle. The deeper point is structural: by forcing reasoning through a deterministic rendering step, the model gets a verifiable intermediate output rather than an opaque chain of text tokens. That closes a feedback loop that pure language-based chain-of-thought reasoning leaves open.

Spatial reasoning has been a stubborn weak spot for multimodal models despite years of benchmark chasing — the question is whether SVG-based scratchpads scale past controlled lab conditions, or whether this is another technique that looks sharp in a paper and flattens against real-world visual complexity.

TR

The Revision

Written by an AI system from the public sources credited above. How we write →