MentalThink gives MLLMs an SVG sketchpad for spatial reasoning

Most efforts to make multimodal models reason better still lean on text chain-of-thought, which is fine when the problem is verbal but leaves geometry and scene questions to fend for themselves. A new paper on arXiv, MentalThink, proposes a different scratchpad: the model writes SVG code, a deterministic renderer draws it, and the model reads the rendered picture back before continuing to reason.

The authors call this a think-with-SVG pipeline, and they train it in two stages. First, supervised fine-tuning aligns the model to SVG syntax. Then multi-turn reinforcement learning teaches it to inspect, revise, and refine its own visual hypotheses across turns. The framing in the abstract is that executable vector graphics act as a verifiable visual workspace, one that lets the model do dynamic perspective taking and compositional scene construction rather than guessing in prose.

On spatial reasoning benchmarks, the paper reports 55.1% on VSIBench and 76.0% on MindCube. Take those as the authors' claim rather than a settled result, because the abstract does not publish the baselines it is measuring against, so the size of any improvement over prior methods is not visible from what has been released.

The honest caveat is that SVG is a schematized, clean-line world. VSIBench and MindCube are spatial reasoning suites, not natural-image chaos, and the abstract does not address how the method holds up when the input scene is messier than a vector sketch. What the paper does not give you at this level of detail is how expensive the multi-turn RL loop is or how much of the gain comes from RL versus SFT.

Still, the direction is interesting. A deterministic renderer is a rare thing in a stack full of stochastic components, and if executable graphics turn into a stable verification loop for spatial reasoning, it is a template robotics, CAD-adjacent, and diagram reasoning teams can borrow.