Google

In language model inference, a common method for generating text is to select the token with the maximum probability from the predicted distribution at each position. This strategy is applied sequentially, where the model's output at each step is determined by the single most likely next token given the preceding sequence.

Greedy Decoding in Language Models

A common text generation strategy involves sequentially selecting the single word with the highest probability at each step to form a sentence. Based on this description, identify one significant advantage and one significant disadvantage of this approach.

Evaluating a Text Generation Strategy

A language model is generating text starting with the prompt 'The ocean is'. Given the model's predicted probabilities for the next two words as detailed below, determine the exact two-word sequence that will be generated if the model exclusively selects the single most likely token at each step. Explain your reasoning for each selection.

Applying a Text Generation Strategy

A language model is generating a two-token sequence. At the first step, it assigns a probability of 0.5 to token 'A' and 0.4 to token 'B'. At the second step, if 'A' was chosen, the model assigns a probability of 0.5 to token 'C'. If 'B' was chosen, it assigns a probability of 0.9 to token 'D'. All other tokens have lower probabilities at each step. Based on this information, which statement accurately analyzes the outcome of a purely sequential, maximum-probability selection strategy?

Learn Before

Related