PlaySuite

A Large-Scale Benchmark for Interactive Visual Intelligence
📄 Paper  ·  💻 Code  ·  ✉️ Contact

🚧 The dataset is coming soon. We are preparing the PlaySuite game collection for public release. Like or follow this dataset to hear when it's available.

What is PlaySuite?

PlaySuite is a benchmark for one question: can multimodal models that understand images and video also act competently over time? Models play real, human-made video games through a standard keyboard and mouse. A video-language judge then watches each episode and rates how far they progressed.

The benchmark covers 5,734 independent games from PyWeek and itch.io. That is more than 100 times as many environments as previous multi-game benchmarks. The games span puzzles, platformers, shooters, strategy, simulation and more, built with Pygame, HTML5, Godot, Unity and many other engines. Because most of these titles sit far from the well-known games that dominate web-scale training data, a model can't succeed by recalling a walkthrough. It has to look, act and adapt.

At a glance

Games in the benchmark	5,734 (itch.io and PyWeek)
Validated launchable games	~17.5K
Original catalog	1.7M titles, filtered for quality, safety and reproducibility
Models evaluated in the paper	14 open VLMs, computer-use agents and VLA models
Progress scale	None → Minimal → Partial → Substantial → Completed

What this dataset will contain

• The PlaySuite game collection. Benchmark-ready environments, with metadata and genre labels for every title.
• Attribution for every game. Links back to each game's original page, so credit stays with its creators.

The companion GitHub repository will host the evaluation harness, the progress judge and the tooling for running PlaySuite at scale.

Why it matters

Our experiments reveal a clear perception–action gap. Current models often understand what is on screen but struggle to turn that into sustained, goal-directed progress. Even the strongest model in our study reaches at least Partial progress in only about one in five runs on itch.io games. PlaySuite is designed to measure progress on exactly this gap, from seeing to doing.

Licensing and attribution

PlaySuite is built entirely on the work of independent game developers, who keep all rights to their games.

• PyWeek games are distributed unmodified under the PyWeek rules, which allow free, non-commercial redistribution. Where a game ships its own license, that license applies.
• itch.io games are restricted to free titles marked for public access. We provide metadata and links to each game's original page, and use the games for non-commercial research and benchmarking only.

Citation

@article{varghese2026playsuite,
  title   = {PlaySuite: A Large-Scale Benchmark for Interactive Visual Intelligence},
  author  = {Varghese, Dheeraj and Vettoruzzo, Anna and Simoncini, Walter and Acevedo Callejas, Michelle Lorena and Derakhshani, Mohammad Mahdi and Meding, Kristof and Vanschoren, Joaquin and Snoek, Cees G. M.},
  journal = {arXiv preprint arXiv:2610.07127},
  year    = {2026}
}

Authors

Dheeraj Varghese¹, Anna Vettoruzzo^2,, Walter Simoncini^1,, Michelle Lorena Acevedo Callejas³, Mohammad Mahdi Derakhshani¹, Kristof Meding^3,†, Joaquin Vanschoren^2,†, Cees G. M. Snoek^1,†

¹University of Amsterdam · ²Eindhoven University of Technology · ³Eberhard Karls Universität Tübingen
^*Equal contribution · ^†Equal advising

Contact: Dheeraj Varghese, d.varghese@uva.nl