Autoregressive LLMs generate text by sampling from estimated probability distributions over the next token, conditional on prior context. We use these probabilities to construct an entropy-based ...