LLMs Generate Narrow Output, Study Finds—Three Methods to Boost Response Diversity

Researchers examining how large language models handle open-ended questions have discovered that AI systems produce far more homogeneous responses than humans do. When given questions with multiple valid answers, LLMs concentrate their outputs around a few popular responses, whereas humans naturally generate a wider range of perspectives and solutions.

The Diversity Gap#

The analysis reveals a stark contrast in response patterns. Human answers to the same questions tend to follow a long-tail distribution—many people give common answers, but a substantial minority offer novel, less-obvious ones. Language models, by comparison, show highly concentrated output, gravitating toward a narrow band of mainstream responses. This suggests that current LLMs may not be tapping into the full spectrum of possible solutions to open-ended problems.

Three Practical Solutions#

The researchers tested three straightforward interventions to improve diversity:

• Temperature sampling: Raising the randomness parameter during generation pushes models away from deterministic, mode-seeking behavior and introduces more variation in outputs.
• Perspective prompting: A single prompt can instruct models to answer from multiple distinct viewpoints, encouraging them to explore different angles on a question without requiring multiple separate queries.
• Model aggregation: Running the same prompt across several different models and combining their outputs naturally yields greater diversity, since different architectures and training data produce different response patterns.

Implications and Limitations#

When combined, these methods substantially increase diversity beyond what a single model produces in isolation. However, even with these approaches, single-model outputs generally remain less diverse than human responses. The authors argue that this work carries important implications for AI policy and governance, particularly for efforts aimed at safeguarding cultural diversity—a cornerstone of democratic participation.