Statement 1 is Correct: At their core, Large Language Models (LLMs) function by analyzing input context and calculating a statistical probability distribution for the next possible words (tokens) in a sequence. In their most basic form of text generation (known as "greedy decoding"), they pick the word with the absolute highest probability. While standard applications often introduce controlled randomness (stochastic sampling like *top-k* or *temperature scaling*) to make text more creative, the foundational mechanism relies on assigning probabilities to select the next word. Statement 2 is Correct: LLM training relies heavily on mathematical optimization. Algorithms such as Stochastic Gradient Descent and backpropagation are used to iteratively adjust billions of model parameters. The explicit goal of this process is to minimize a "loss function," which mathematically represents the model's prediction errors compared to the actual training data. Statement 3 is Incorrect: LLMs do not produce inherently unbiased outputs. Because they learn from massive, human-generated datasets scraped from the internet, they are well-documented to inherit, reflect, and sometimes amplify historical and societal biases (e.g., gender, racial, or cultural stereotypes) present in their training data. Therefore, the correct option is B.