Alice: A System for Online Executable World Model Learning Under Prior Misalignment

A recent preprint on arXiv presents Alice, a closed-loop framework designed for online executable world-model learning in situations where prior knowledge is not aligned. The system interprets unsuccessful candidate updates as structural indicators: when a candidate accounts for a new transition but fails to retain previously explained ones, the resulting preservation conflict uncovers dynamics that were previously mixed by the current program. Alice transforms these conflicts into hypothesis classes that yield compact, class-stratified preservation counterexamples for updates, facilitating exploration towards new and less represented transitions. This research tackles the problem of deriving state-dependent dynamics solely from interaction data, without relying on rule descriptions, reward signals, or reliable lexical priors.