Growing a Tail: Increasing Output Diversity in Large Language Models

Michal Shur-Ofry, Bar Horowitz-Amsalem, Adir Rahamim, Yonatan Belinkov

Abstract

How diverse are the outputs of large language models when diversity is desired? We examine the diversity of responses of various models to questions with multiple possible answers, comparing them with human responses. Our findings suggest that models' outputs are highly concentrated, reflecting a narrow, mainstream 'worldview', in comparison to humans, whose responses exhibit a much longer-tail. We examine three ways to increase models' output diversity: 1) increasing generation randomness via temperature sampling; 2) prompting models to answer from diverse perspectives; 3) aggregating outputs from several models. A combination of these measures significantly increases models' output diversity, reaching that of humans. We discuss implications of these findings for AI policy that wishes to preserve cultural diversity, an essential building block of a democratic social fabric.

Introduction: The Problem of Algorithmic Monoculture

As Large Language Models (LLMs) increasingly serve as primary gateways to information, a significant risk of cultural homogenization emerges. LLMs are fundamentally trained to predict the most statistically probable next token, a design that naturally favors the "mean"—the most popular, mainstream, and well-known concepts. This phenomenon, termed "algorithmic monoculture," can lead to a narrow "worldview" where users are repeatedly exposed to the same limited set of perspectives.

This research investigates "exposure diversity": the extent to which users encounter a broad range of content when asking open-ended questions. While previous studies have explored algorithmic monoculture in decision-making contexts (like credit scoring), this paper focuses on the social and democratic implications of LLMs providing a homogenized view of history, culture, and geography.

Methodology: Measuring and Promoting Diversity

The researchers evaluated eight prominent LLMs, including GPT-4, Gemini, and Claude, through a three-stage experimental design:

  1. Baseline Evaluation: Models were prompted with open-ended questions (e.g., "List three cities worth visiting") at a standard low temperature (0.3).
  2. Diversity-Promoting Treatments: The study tested individual and combined interventions, including:
  3. Temperature Scaling: Increasing the temperature hyperparameter to 1.0 to introduce more randomness.
  4. Diversity Prompting: Explicitly instructing models to "Answer from diverse perspectives."
  5. Model Aggregation: The researchers aggregated outputs from six different API-based models to determine if their collective "worldviews" provided a broader distribution than any single model.

Diversity was quantified using Shannon Entropy and unique item counts. To establish a "gold standard," the researchers collected human responses via a thread on "X" (formerly Twitter), providing a baseline for the "long-tail" distributions typical of human thought.

Key Results: Bridging the Diversity Gap

The baseline results revealed a stark contrast between human and machine outputs. In the culture domain (TV series), humans provided a wide array of unique suggestions, whereas LLMs tended to converge on a few "prestige" titles.

Figure 1: Baseline Stage Diversity across History, TV Series, and Cities, comparing GPT-3.5 and Gemini against the human "Twitter Feed" baseline.

As shown in Figure 1, the "short tail" of LLM responses is evident. For 19th-century figures, models repeatedly generated names like Charles Darwin and Karl Marx, while humans offered a much broader variety.

Synergistic Effects and Aggregation

A key finding was the synergy between temperature increases and diversity prompting. The improvement in diversity from combining these treatments was greater than the sum of their individual effects. However, true human-level diversity was only achieved through model aggregation. By combining six different models using diversity-inducing treatments, the researchers generated 58 to 60 unique TV series, actually exceeding the diversity of the 30 human responses (25 unique series).

Figure 2: The "long tail" of human responses for TV series collected from a Twitter feed, showing a high number of unique items with low individual frequency.

Importantly, this increase in diversity did not come at the cost of accuracy. Error rates remained consistently low (0-2%), suggesting that models possess "hidden" diversity that can be unlocked without inducing hallucinations.

Discussion: Implications for AI Policy

The study demonstrates that while LLMs default to a narrow, Western-centric worldview, they contain the latent capacity for significant diversity. The researchers suggest that the "short tail" is a result of training objectives rather than a lack of knowledge.

These findings have profound implications for AI policy. To preserve the cultural diversity essential for a democratic social fabric, policymakers should encourage AI literacy and foster market competition. Ensuring that users are not trapped in an algorithmic monoculture requires both technical interventions—like those demonstrated in this study—and a broader commitment to maintaining a diverse information ecosystem.

Summary updated

Abstract

How diverse are the outputs of large language models when diversity is desired? We examine the diversity of responses of various models to questions with multiple possible answers, comparing them with human responses. Our findings suggest that models' outputs are highly concentrated, reflecting a narrow, mainstream 'worldview', in comparison to humans, whose responses exhibit a much longer-tail. We examine three ways to increase models' output diversity: 1) increasing generation randomness via temperature sampling; 2) prompting models to answer from diverse perspectives; 3) aggregating outputs from several models. A combination of these measures significantly increases models' output diversity, reaching that of humans. We discuss implications of these findings for AI policy that wishes to preserve cultural diversity, an essential building block of a democratic social fabric.
BibTeX
@misc{shurofry2024growing,
  author = {Michal Shur-Ofry and Bar Horowitz-Amsalem and Adir Rahamim and Yonatan Belinkov},
  title = {{Growing a Tail: Increasing Output Diversity in Large Language Models}},
  year = {2024},
  eprint = {2411.02989},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2411.02989},
}
Figure