A new study published on arXiv analyzes the grammatical and lexical diversity of large language models (LLMs) compared to human-written text. Researchers found that newer, instruction-tuned LLMs exhibit reduced syntactic and lexical diversity when compared to older models and human-authored news articles from The New York Times. This suggests that while instruction tuning improves coherence, it may also narrow the expressive range of LLM outputs. AI
Summary written by gemini-2.5-flash-lite from 3 sources. How we write summaries →
IMPACT Suggests instruction tuning may reduce LLM expressive range, impacting creative and nuanced text generation.
RANK_REASON Academic paper analyzing LLM output characteristics.
