Code-as-World
mirros's release for Code as Worlds: Agentic Discovery of Executable World Representations for Physical Reasoning — Python, Apache-2.0, 198 stars and 1 fork, repository created 2026-08-27 and last pushed the following morning. What it ships is inference and evaluation code, not training code: the two code-as-world-vl checkpoints (4B and 9B) on HuggingFace, a quantiphy evaluation harness, a vLLM serving recipe, and one bundled MuJoCo example.
T1 for what it is, and only that. This page grades the repository as the artifact's own facts — the code, the checkpoints, the license, what runs. Every performance number lives in the companion code-as-world-report, graded T3, because MirroS ran that benchmark on models MirroS trained against baselines MirroS configured.
The claim in the README
Pixels are evidence of the physical world, not its ontology.
The stated problem is that a pixel record says how a scene looked from one viewpoint at one moment and does not say what is in it, how it is structured, or what governs its change. The proposed answer is to represent the world as executable code — an executable-world-representation — and to recover that code from observations with an agentic-discovery-loop, described in the README as turning raw observation into "explicit states, dynamics, and mechanisms that capture not only what was seen, but the underlying world that could have produced it."
The phrase that matters to this wiki is the one after it: "This provides scalable physical supervision." The repository's subject is world representation; its use, and the only part it measures, is manufacturing training labels.
What is actually in the box
| Piece | What it does |
|---|---|
code_as_world.evaluation |
Runs 4B or 9B over the QuantiPhy validation CSV, writes evaluator-compatible predictions |
code_as_world.simulation |
The one worked example — a "ballistic soccer case" rendered in MuJoCo |
| vLLM recipe | vllm serve with a Qwen3.5 no-think chat template, 16 frames per video, 4,608-token context |
| Checkpoints | Code-as-World-VL-4B, Code-as-World-VL-9B on HuggingFace |
Both TODO items are checked off, so the release is complete as scoped. The discovery agent itself is not in the repository. The loop that produces executable worlds — the thing the paper is about — ships as a description and one MuJoCo demo; what you can run is the model trained on its output.
Reading the dependency list
The acknowledgements name the stack the loop is assembled from, and it is worth recording that none of it is MirroS's: MuJoCo for simulation, SAM 3 / DA3 / VGGT-Omega / SAM 3D for perception and reconstruction, Wan and VACE for video generation, quantiphy for evaluation. The contribution is the loop and the supervision it produces, not a new perception or simulation component.
Provenance note
The BibTeX entry cites arXiv preprint arXiv:2608.xxxxx — a placeholder. As of ingest the report is
self-published at mirros.ai and carries no arXiv identifier, no venue and no named authors beyond
"MirroS Team". Not a flaw in the work; a fact about what can be checked.
Related
- code-as-world-report — the technical report, where the numbers and the limitations are
- code-as-world-vl — the model family this repository serves
- executable-world-representation · agentic-discovery-loop · quantiphy
- training-data-quality — the corpus's other answer to where labels come from