Patterns of Priming in Production: Lexical, Semantic and Structural Alignment in Language Model Generation

On September 7, 2026, arXiv cs.AI published a study on the generation process of language models (LM). Through a series of controlled experiments on preposition structure completion, the study confirmed that LM exhibits a structural priming effect during generation. It was found that when the semantics of sentences are coherent, LM is particularly sensitive to structural priming; among all structures, those with double objects show a greater relative increase in priming intensity, while those with frequent preposition-object structures show a greater absolute increase. Additionally, structural priming is not only influenced by lexical-semantic coherence but also results in higher levels of lexical-semantic repetition in the generated completions. Overall evidence suggests that structural priming in language models operates at multiple linguistic representation levels and is facilitated and constrained by syntactic, lexical, and semantic alignment.