Joint Learning of Hierarchical Neural Options and Abstract World Model
Wasu Top Piriyakulkij, Wolfgang Lehrach, Kevin Ellis, Kevin Murphy · 2026
Read Joint Learning of Hierarchical Neural Options and Abstract World Model by Wasu Top Piriyakulkij, Wolfgang Lehrach, Kevin Ellis, Kevin Murphy (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-02799.txt.
Description
Wasu Top Piriyakulkij and 3 co-authors teach agents composable skills and a world model at the same time, so each makes the other cheaper to learn, in a 2026 paper. Building new skills by composing old ones is a long-standing goal. Model-free hierarchical reinforcement learning can do it, but only with large amounts of data.
Their agent jointly learns an abstract world model, compressed across both states and time, and a set of hierarchical neural options, reusable skill units. Learning proceeds sample-efficiently because the model guides skill acquisition while skills structure the model. On a subset of object-centric Atari games, the method learned more skills from less data than baselines. It also showed learning and generalization capabilities the baselines lacked entirely. The authors frame joint learning, not bigger hierarchies, as the route to agents that accumulate skills like people do.
Catalog record
| CREATOR | Wasu Top Piriyakulkij, Wolfgang Lehrach, Kevin Ellis, Kevin Murphy |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 56 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-02799