Compute the Distribution of LLM Decoding Stopping Time and Build a Strategy Against Adversaries

Problem: Analyze LLM Generation Termination and Choose a Minimax Defense Consider a simplified language-model generation loop. At every decoding position, the model can produce a distinguished token, ending the generation. Your task has two parts. Part A: Distribution of the Termination Step Given an upper bound N on generated tokens (generation cannot continue beyond N; a run that has not ended beforehand terminates at N) together with per-position stopping probabilities:…