Question 33
IWhich property is most characteristic of a large language model (LLM) compared with earlier NLP systems?
Correct answer: B
Explanation
Large language models are trained end-to-end on massive text corpora with the objective of predicting the next token. That training lets them learn broad language patterns and “generalize to many tasks via prompting,” unlike earlier NLP systems that relied more on task-specific rules or features.
Why each option is right or wrong
A. It relies exclusively on manually designed grammar rules for each language.
Rule-based grammar systems are earlier symbolic NLP approaches, not the defining pattern of LLMs.
B. It is trained end-to-end on massive text corpora to predict the next token, enabling it to generalize to many tasks via prompting.
LLMs are typically built by pretraining a single neural model on very large text datasets with the self-supervised objective of next-token prediction, rather than hand-engineered rules or narrowly supervised task pipelines. That architecture is what allows the same model to be adapted to many downstream tasks through prompting, because the model has already learned broad statistical patterns in language from the corpus rather than being trained separately for each task.
C. It can only perform the single task it was explicitly fine-tuned for.
LLMs are general-purpose models; fine-tuning can specialize them, but prompting alone supports many tasks.
D. It requires labeled intent data for every downstream use case.
LLMs often learn from unlabeled text, so every downstream use case does not require labeled intent data.