Why do the big LLMs use correlative conjunctions all the time?

Understanding the Frequent Use of Correlative Conjunctions in Large Language Models

Have you ever noticed how large language models (LLMs) tend to frequently employ correlative conjunctions—phrases like “not only… but also” or “both… and”—in their generated texts? This pattern raises an intriguing question: what underlying factors drive these models to consistently select such constructions, especially in their truncated forms like “not just… but also”?

From a technical perspective, the utilization of correlative conjunctions by LLMs can be attributed to several interconnected aspects of their training and architecture. These models learn language patterns based on vast datasets that often contain sophisticated and varied sentence structures. Consequently, they develop a propensity for using constructions that are both grammatically correct and stylistically effective in emphasizing the relationship between ideas.

Specifically, the fragment “not just… but also” exemplifies a common pattern for highlighting additional importance or depth in a statement. For instance, a sentence such as “Fishing is not just about catching the biggest fish—it’s about experiencing nature,” effectively underscores the multifaceted nature of the activity. The model’s frequent choice of this structure likely stems from its statistical prominence within the training corpora, as well as its clarity in conveying nuanced ideas.

While the exact “reason” for this tendency remains rooted in the model’s learned language probabilities, ongoing research continues to explore whether these patterns are simply a product of training data distribution or if they reflect deeper syntactic and semantic preferences ingrained in human language usage. Understanding these tendencies not only sheds light on the inner workings of LLMs but also guides the development of more nuanced and context-aware language generation systems.

In summary, large language models often favor correlative conjunctions because these structures are prevalent, expressive, and stylistically effective within their training data. Recognizing this pattern enhances our appreciation of how LLMs mimic human language nuances and informs future improvements in natural language processing.