First Token Cutoff LLM sampling

From a theoretical standpoint, the best reply provided by an LLM is obtained by always picking the token associated with the highest probability. This approach makes the LLM output deterministic,...