What does a generative language model do when it produces text?
Why B is the answer
It predicts the next token repeatedly, based on patterns it learned
The model generates output one token at a time, based on patterns from training.
In detail
- A language model predicts the next token repeatedly.
- It works from patterns learned in training.
- It does not look answers up, which is why it can be wrong.
Use case
An illustrative scenario, not a real incident.
Asked to finish "The capital of France is", a model predicts "Paris" from learned patterns. It repeats that token by token.
See it in different ways
flowchart LR A["Prompt"] --> B["Split into tokens"] B --> C["Model predicts the next token"] C --> D["Add it to the text"] D --> C D --> E["Response"]
stateDiagram-v2 [*] --> Tokens: prompt becomes tokens Tokens --> Predict Predict --> Append: pick the next token Append --> Predict: not finished Append --> [*]: stop or token limit
flowchart LR S0["Prompt"] --> S1["Predict next token"] --> S2["Add it"] --> S3["Repeat until done"]
Key point: Models predict the next token repeatedly.
Why the other options are wrong
- A. It looks up each full answer in a stored database of responses
- It does not look answers up in a stored database.
- C. It copies the closest matching document from the internet word for word
- It does not copy a document from the internet.
- D. It runs a fixed set of rules written by hand for each topic
- It is not a set of written rules.