README.md

---
title: Recursive Play
emoji: 🧩
colorFrom: green
colorTo: yellow
sdk: static
app_file: index.html
pinned: false
datasets:
- EnvLoop/Recursive-Play-Bench
tags:
- benchmark
- evaluation
- puzzles
- rl-environment
short_description: Play real puzzle rules and explore 11 model configurations.
---

# Recursive Play

By **EnvLoop Research** — research@envloop.ai

A browser-only practice demo and results explorer for Recursive-Play: 150 authored candidate environments, six levels each, with 1,650 independently eligible outcomes from 11 configurations.

- **Try four real environments.** The existing deterministic TypeScript rules run locally in the browser. Reset restores the exact initial instance and costs one action. A practice session allows 512 actions and 720 seconds across the six sequential levels. New practice sessions are separate from the published model results. Neutral numbered controls preserve the puzzle interface; button 6 can also receive a click on the pixel grid.
- **Explore verified results.** The rankings, four vector plots, selected actual episode frames and 24-page paper reuse the independently reviewed V14 artifacts. No new model, game-server or training requests are made by this Space.
- **Practice is source-exposed.** This static demo is not a secure blind-evaluation environment and does not submit scores to the benchmark. Solver results in the paper were collected separately with withheld authored rules and constructive witnesses.

[Research collection](https://huggingface.co/collections/EnvLoop/recursive-play) · [Dataset and benchmark artifacts](https://huggingface.co/datasets/EnvLoop/Recursive-Play-Bench) · [Paper preview](https://huggingface.co/datasets/EnvLoop/Recursive-Play-Bench/blob/cb12d923f369109626b083ae6e44a3139abedca9/recursive-play-paper.pdf) · [Reproducible paper sources](https://huggingface.co/datasets/EnvLoop/Recursive-Play-Bench/blob/cb12d923f369109626b083ae6e44a3139abedca9/paper-source.zip)

The work measures task construction and solving outcomes. The illustrated learning loop is a proposed future experiment; training gains or recursive improvement have not been measured here. All 150 environments remain authored candidates: quality flags and the limits of difficulty calibration are documented in the paper and dataset.

## Explore the benchmark on Hugging Face

The dataset provides six views in the native Dataset Viewer. Start with the model summary to compare complete-task success and partial level progress, then inspect the individual outcomes or task records. Hugging Face supports column sorting, search and filters; numeric columns in the summary views also support distributions. The SQL console offers a way to examine the tables in the browser.

| View | Rows | What it shows |
|---|---:|---|
| [Models](https://huggingface.co/datasets/EnvLoop/Recursive-Play-Bench/viewer/models/test) | 11 | One summary for each evaluated configuration, including tied ranks and the 150-task and 900-level denominators. |
| [Tasks](https://huggingface.co/datasets/EnvLoop/Recursive-Play-Bench/viewer/tasks/test) | 150 | The canonical environments, seeds and curriculum metadata. |
| [Results](https://huggingface.co/datasets/EnvLoop/Recursive-Play-Bench/viewer/results/test) | 1,650 | The independently eligible episode records underlying the paper. |
| [Challenges](https://huggingface.co/datasets/EnvLoop/Recursive-Play-Bench/viewer/challenges/test) | 15 | Tasks that none of the eleven tested configurations fully completed under the stated budget. |
| [Families](https://huggingface.co/datasets/EnvLoop/Recursive-Play-Bench/viewer/families/test) | 110 | Progress grouped by configuration and task family. |
| [Completion depths](https://huggingface.co/datasets/EnvLoop/Recursive-Play-Bench/viewer/depths/test) | 77 | Counts of episodes completing zero through six levels for each configuration. |

The summary, family and depth views are derived from the same 1,650 results. They are not additional trials, and their row counts should not be added to the episode total. The challenge subset records observed non-completion; it does not certify that the tasks are intrinsically hard or that future agents cannot solve them.

The [Recursive Play collection](https://huggingface.co/collections/EnvLoop/recursive-play) connects the paper, dataset and practice Space. The linked paper and source archive above are pinned to the published V14 dataset commit. Dataset releases can be reproduced by selecting a commit under **Files and versions** and passing its hash as the `revision` when loading data. The dataset card documents the schemas and interpretation of missing values.

## Files and reproduction

`index.html`, `styles.css` and prebuilt `app.js` need no server-side build. `app-source.ts` and the four `engine-*.ts` files expose the local practice implementation. `benchmark.json` holds selected display data; `asset-provenance.json` records source hashes. The PDF, SVG figures and episode PNGs are unmodified copies of the reviewed artifacts.

All files use relative paths. No external fonts, analytics, paid APIs or user credentials are required. Requests concerning source permissions or reuse: research@envloop.ai. No project-wide license is asserted by this card.

## Citation

```bibtex
@misc{envloop_recursive_play,
  title = {Recursive-Play: AI-Generated Interactive Challenges for Improving AI Agents},
  author = {{EnvLoop Research}},
  url = {https://huggingface.co/datasets/EnvLoop/Recursive-Play-Bench}
}
```