Synthetic Physics as Supervision: Learning Real-World Physical Reasoning in Vision-Language Models

Vision-language models (VLMs) remain unreliable on visually grounded physics reasoning in real-world media, and human-labelled physics supervision is expensive. We study whether synthetic simulator scenes alone can supply an effective training signal. We fine-tune Qwen3-VL-30B on 14,597 rigid-body and fluid scenes with free-text answers read directly from simulator state, using no human annotation and no local GPU. On PhysBench Test (n=9,786), this synthetic-only SFT lifts accuracy from 40.7% to 47.6% (+6.9pp), improves 27/39 subtasks, and yields a +20.0pp gain on the Scene domain. A follow-up GRPO stage with simulator-verifiable rewards preserves this aggregate gain (+7.0pp over baseline) while adding targeted improvements on the un-trained general:relationships domain, and a data-fidelity probe at both stages shows the gains track the physical quality of the synthetic signal. On a vision-essential ScienceQA-Physics probe, prior textbook-physics ability is not significantly degraded, so the model remains a generalist VLM. The full pipeline costs under \$30 in hosted-API credits, providing evidence that synthetic physics is a practical and reproducible supervision signal for grounded visual reasoning in VLMs.