Guardat en:
| Autor principal: | |
|---|---|
| Format: | Recurso digital |
| Idioma: | |
| Publicat: |
Zenodo
2026
|
| Matèries: | |
| Accés en línia: | https://doi.org/10.5281/zenodo.20429921 |
| Etiquetes: |
Afegir etiqueta
Sense etiquetes, Sigues el primer a etiquetar aquest registre!
|
Taula de continguts:
- <p>First public release of nanoAWM: a tiny, from-scratch (pure-Python, no PyTorch/NumPy) action-conditioned world model for tool-using agents, plus the MiniOS task suite it learns on.</p> <p>Honest headline: on a genuinely held-out split with object and action vocabulary disjoint from training, the learned world-model planner scores 0.524, versus 0.067 for the best non-world-model baseline and 1.000 for an oracle upper bound. The in-distribution 1.000 is a sanity check, not the result.</p> <p>Documented negative results are included by design (action-paraphrase collapse, cross-surface OOD, lexicon holdout). Code, the trained model, the dataset, the paper, and all audits are in the repository.</p>