Game2Real-Bench: Do Toy Games Reveal and Transfer to Real-World Visual Reasoning?
Qingan Liu, Juntong Feng, Yue Zhu, Yujie Cheng, Xinzhe Han, Shuai Hao, Yunzhi Zhuge, Haiwen Diao and 1 others
Dalian University of Technology · Nanyang Technological University
Published October 13, 2026 · on the feed October 19, 2026 · ▲ 5 · arXiv 2610.18834
The abstract
Visual reasoning is an important capability for vision-language models (VLMs) to understand visual evidence, spatial structure, constraints, and state changes. Existing evaluations mainly rely on real-world data or procedurally generated tasks. Real-world data preserve natural complexity, while procedural games provide controllable generation and verifiable solutions. However, it remains unclear whether abilities measured in toy games correspond to real-world visual reasoning and whether game examples can support transfer. We introduce Game2Real-Bench, a benchmark that studies these connections across four capability groups, including fine-grained perception, spatial reasoning, constraint reasoning, and sequential planning. We construct 1,000 procedurally generated game samples with executable verification and curate 400 real-world samples aligned by shared reasoning requirements rather than visual similarity. We evaluate 25 representative VLMs and observe strong game-to-real consistency, with Pearson and Spearman correlations of 0.935. We further study one-shot game-to-real transfer on four models. Constraint reasoning improves for several models, while fine-grained perception declines with game demonstrations. We also construct Game2Real-Dataset with nearly 24K verified game samples across 18 tasks and 65 QA families. Our benchmark, dataset, and code are available at https://github.com/QingAnLiu/Game2Real-Bench.
The Brief
2:23A short podcast summary of the paper. Written and voiced by AI.
Transcript
Game2Real-Bench. Do Toy Games Reveal and Transfer to Real-World Visual Reasoning?. By Qingan Liu and eight others, from the Dalian University of Technology and Nanyang Technological University. Published on the thirteenth of October. Its code repository is worked out: about five hundred and fifty source files, with a readme and a licence.
Vision-language models need visual reasoning to make sense of visual evidence, spatial structure, constraints and changes of state. Evaluations of that ability come in two kinds. Real-world data keeps the natural complexity of actual scenes. Procedurally generated games offer controllable generation and solutions that can be verified. Each has something the other lacks. What nobody has settled is whether the two agree. Does doing well on a toy game say anything about reasoning in the real world? And can game examples actually help a model once it faces real images?
Game2Real-Bench is built to study both of those connections, the diagnostic one and the transfer one. It organises visual reasoning into four capability groups: fine-grained perception, spatial reasoning, constraint reasoning, and sequential planning. Each question can then be asked group by group, rather than for visual reasoning as a single lump.
On the game side there are one thousand procedurally generated samples, each with executable verification. On the real side there are four hundred curated samples. The pairing between the two is the key design choice. Real samples are aligned with game samples by the reasoning they require, not by how they look. A game and a photograph can share nothing visually and still demand the same kind of thinking.
Across twenty-five representative models, game and real performance track each other closely, with Pearson and Spearman correlations of zero point nine three five. The games, in other words, do reveal something about real-world ability.
Transfer is a more mixed story. In one-shot game-to-real tests on four models, constraint reasoning improves for several of them, while fine-grained perception actually declines when game demonstrations are shown. Alongside the benchmark, the authors release Game2Real-Dataset, with nearly twenty-four thousand verified game samples covering eighteen tasks and sixty-five question answering families.
The title asks two questions, and the answers diverge. As a mirror of real-world reasoning, toy games hold up well. As a teacher, they help with some skills and hurt others, and perception is where the help turns into harm.
More with an account
Listen to every day’s papers
Daily playlists, deep dives, quests for the papers, and quizzes on their key terms.
Join free