Learning POMDP World Models from Observations with Language-Model Priors

Researchers introduced Pinductor, a method that uses language-model priors to learn POMDP world models from limited observation-action data.

Pinductor asks an LLM to propose and refine candidate hidden-state models, then fits them using belief-based likelihood.

In experiments, it matched methods with access to hidden states and beat tabular baselines on sample efficiency.

The result suggests pretrained language models can help agents build useful internal models of uncertain environments with less interaction.