New probabilistic language tries framework unifies AI model compression, decision-making, and computational efficiency through structured prefix representation

Probabilistic language tries (PLTs) make explicit the prefix structure in generative models by assigning conditional token probabilities to each edge

PLTs function as optimal lossless compressors via frequency-weighted interval encoding, extending arithmetic coding to model-conditioned distributions

The framework serves as a policy representation for sequential decision problems including games, search algorithms, and robotic control tasks

PLTs enable memoization indexing that replaces full model execution with structured retrieval for repeated inference queries

A prior-guided caching theorem demonstrates that PLT-guided caches achieve lower expected inference costs than empirical-frequency caches below certain query thresholds