“In autoregressive LLMs, the probability of an answer remaining factually correct decreases exponentially with the number of tokens generated.”