Learning POMDP World Models from Observations with Language-Model Priors
Valentin Six 1,* , Frederik Panse 1,4,* , Mathis Fajeau * , Lancelot Da Costa 1,* Mridul Sharma 2,† , Alfonso Amayuelas 3,† , Tim Z. Xiao 1,4 , David Hyland 5 Philipp Hennig 4 , Bernhard Schölkopf 1,6 1 Max Planck Institute for Intelligent Systems 2 IRIIS 3 University of California, Santa Barbara 4 University of Tübingen 5 University of Oxford 6 ELLIS Institute Tübingen * Equal first-author contribution. † Equal second-author contribution. · 2026
Read Learning POMDP World Models from Observations with Language-Model Priors by Valentin Six 1,* , Frederik Panse 1,4,* , Mathis Fajeau * , Lancelot Da Costa 1,* Mridul Sharma 2,† , Alfonso Amayuelas 3,† , Tim Z. Xiao 1,4 , David Hyland 5 Philipp Hennig 4 , Bernhard Schölkopf 1,6 1 Max Planck Institute for Intelligent Systems 2 IRIIS 3 University of California, Santa Barbara 4 University of Tübingen 5 University of Oxford 6 ELLIS Institute Tübingen * Equal first-author contribution. † Equal second-author contribution. (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-2605-13740.txt.
Description
Valentin Six and nine colleagues introduce Pinductor, a world-model learner in which a word model offers trial world models from a few observation-action trajectories. It step by step refines them to raise a belief-based fit score. Despite using strictly less data, Pinductor matches the scores and thrift of word-model-based POMDP learning methods given private access to the hidden state. So, these results place word-model priors as a handy tool for sample-thrifty world-model learning under partial seeing.
Catalog record
| CREATOR | Valentin Six 1,* , Frederik Panse 1,4,* , Mathis Fajeau * , Lancelot Da Costa 1,* Mridul Sharma 2,† , Alfonso Amayuelas 3,† , Tim Z. Xiao 1,4 , David Hyland 5 Philipp Hennig 4 , Bernhard Schölkopf 1,6 1 Max Planck Institute for Intelligent Systems 2 IRIIS 3 University of California, Santa Barbara 4 University of Tübingen 5 University of Oxford 6 ELLIS Institute Tübingen * Equal first-author contribution. † Equal second-author contribution. |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 172108 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.13740