Efficient Guided Generation for Large Language Models

Formalizes text generation as transitions between the states of a finite-state machine, allowing an index to be constructed over the model's vocabulary so that grammar-constrained decoding becomes a per-token mask rather than per-token generation-and-test. The abstract's framing - an index over the vocabulary, transitions between FSM states - is the basis for the grammar-mask flop record and for why the cost scales with vocabulary size, not with model size.