PlaySuite
PlaySuite: A Large-Scale Benchmark for Interactive Visual Intelligence
PlaySuite evaluates interactive visual intelligence across 5,734 diverse video games using a closed-loop agent harness and Video-LLM-as-a-judge to measure goal-directed performance.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info playsuite
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
12 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Qwen3-Omni-30B-A3B-Thinking | 0.8 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| Qwen2.5-VL-7B | 0.8 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| google/gemma-4-26B-A4B-it | 0.7 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| UI-TARS-1.5-7B | 0.7 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| Qwen3-VL-8B-Instruct | 0.7 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| EvoCUA-8B | 0.7 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| Qwen2.5-Omni-7B | 0.7 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| google/gemma-4-E4B-it | 0.7 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| OpenCUA-7B | 0.6 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| GUI-Owl-1.5-8B | 0.6 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| InternVL2.5-8B | 0.5 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| OS-Atlas-Pro-7B | 0.5 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsAbout
PlaySuite comprises 5,734 executable video games curated from itch.io and PyWeek, spanning engines like Unity and Godot. It uses a unified HPC-optimized interaction framework allowing agents to control games via keyboard and mouse inputs.
Scoring relies on a Video-LLM-as-a-judge protocol that maps observable gameplay milestones to standardized progress levels. Results indicate a perception-action gap, with current models struggling in spatial grounding and sustained interaction despite strong reasoning capabilities.