# Schema: Discovering Unknown Environments via Agentic Program Induction

> Infers executable models of game state and transition rules from interaction. Agents revise these models against recorded history, search within them, and compare predicted observations with real outcomes while executing plans.

- Authors: Guanning Zeng; Jiani Wang; Wenjie Ma; Shaofeng Yin; Chenyang Wang; Shichen Liu; Angjoo Kanazawa; Wode Ni; Xiuyu Li; Andrea Zanette; Haiwen Feng
- Venue: arXiv
- Status: Preprint
- Publication / posting date: September 30, 2026
- Date source: arXiv v1 first-submission date; the July project announcement and October repository revision are separate.
- First arXiv submission: September 30, 2026
- arXiv: 2609.39140
- Reviewed arXiv version: v1
- DOI: 10.48550/arxiv.2609.39140
- Sources reviewed: 2026-10-20
- Added to the catalog: 2026-10-20
- Original paper: <https://arxiv.org/abs/2609.39140>
- PDF: <https://arxiv.org/pdf/2609.39140v1>
- Code: <https://github.com/guanning03/Schema>
- Project: <https://guanning03.github.io/schema-webpage/>
- BibTeX: <https://awesomerea.org/papers/citations/paper-schema-program-induction.bib>
- Catalog page: <https://awesomerea.org/papers/paper-schema-program-induction/>
- Record JSON: <https://awesomerea.org/papers/paper-schema-program-induction/record.json>
- Chinese version: [简体中文](https://awesomerea.org/zh/papers/paper-schema-program-induction/index.md)

## Scope and limitations

Evaluation uses 25 public ARC-AGI-3 games, 21 public DiG-bench games and MazeBench. Matching recorded transitions establishes history consistency, not original-source recovery or correctness on unseen states. RHAE measures completion and environment-action efficiency, not elapsed time or total computation. Component ablations cover ten ARC games; MazeBench compares recorded trajectories with a leaderboard median rather than a controlled human experiment. Transfer to arbitrary software or stochastic environments is not demonstrated.

## Why it fits this collection

Provides a concrete observation-to-program-and-replay loop for black-box behavioral reconstruction, while distinguishing a useful predictive model from a complete game clone.

## Publication and review notes

The eleven-author v1 list and title match arXiv and the current project. DataCite registered the arXiv DOI on 2026-10-01; the article uses arXiv’s non-exclusive distribution license. The author-linked MIT code was inspected at 84c075fadfe937b97a3d1ad2acb42d1233f55ffe. Benchmark engines/services retain separate requirements. The linked trace dataset README at 0a0858a9e61a68d1e83142d9e4e3d66fc779c403 describes 50 July ARC trajectories, including a mixed-model collection. No trace-data license was verified.

## Primary sources

1. <https://arxiv.org/abs/2609.39140>
2. <https://arxiv.org/html/2609.39140v1>
3. <https://api.datacite.org/dois/10.48550/arxiv.2609.39140>
4. <https://schema-harness.github.io/>
5. <https://guanning03.github.io/schema-webpage/>
6. <https://github.com/guanning03/Schema/blob/84c075fadfe937b97a3d1ad2acb42d1233f55ffe/README.md>
7. <https://github.com/guanning03/Schema/blob/84c075fadfe937b97a3d1ad2acb42d1233f55ffe/LICENSE>
8. <https://huggingface.co/datasets/schema-harness/arc-agi-3-schema-traces/blob/0a0858a9e61a68d1e83142d9e4e3d66fc779c403/README.md>

## About this entry

Reviewed against the paper and its primary sources. Cite the original authors using the linked paper or BibTeX.
Original editorial content: MIT. Linked projects, papers and assets retain their own rights and licenses.