When LLMs Play the Telephone Game: Cultural Attractors as Conceptual Tools to Evaluate LLMs in Multi-turn Settings (2407.04503v3)

Abstract: As LLMs start interacting with each other and generating an increasing amount of text online, it becomes crucial to better understand how information is transformed as it passes from one LLM to the next. While significant research has examined individual LLM behaviors, existing studies have largely overlooked the collective behaviors and information distortions arising from iterated LLM interactions. Small biases, negligible at the single output level, risk being amplified in iterated interactions, potentially leading the content to evolve towards attractor states. In a series of telephone game experiments, we apply a transmission chain design borrowed from the human cultural evolution literature: LLM agents iteratively receive, produce, and transmit texts from the previous to the next agent in the chain. By tracking the evolution of text toxicity, positivity, difficulty, and length across transmission chains, we uncover the existence of biases and attractors, and study their dependence on the initial text, the instructions, LLM, and model size. For instance, we find that more open-ended instructions lead to stronger attraction effects compared to more constrained tasks. We also find that different text properties display different sensitivity to attraction effects, with toxicity leading to stronger attractors than length. These findings highlight the importance of accounting for multi-step transmission dynamics and represent a first step towards a more comprehensive understanding of LLM cultural dynamics.

• The paper introduces a transmission chain framework to analyze cumulative text transformations, revealing attractor states in LLM outputs.
• The paper finds that open-ended tasks yield stronger attractor effects than constrained rephrasing, influencing text convergence.
• The paper compares different LLMs to show how initial inputs and model choices drive varying semantic drift and bias evolution.

Cumulative Changes and Attractors in Iterated Cultural Transmissions of LLMs

The paper "When LLMs Play the Telephone Game: Cumulative Changes and Attractors in Iterated Cultural Transmissions" explores the understanding of how information evolves when transmitted repeatedly through LLMs. Specifically, the paper adopts a transmission chain design from human cultural evolution literature, aiming to uncover how LLM-generated texts change across multiple iterations and whether these changes reveal certain attractor states.

Research Objectives

The primary objective of this paper is to examine the collective behaviors and information distortions that arise from iterated interactions between LLMs. This is motivated by the growing presence of LLM-generated content online, necessitating an understanding of the transformation and potential biases that occur in multi-turn interactions. The paper systematically explores how properties such as toxicity, positivity, difficulty, and length evolve in texts through these iterative transmissions.

Methodology

The paper employs a transmission chain design where LLM agents iteratively receive, produce, and transmit texts across a sequence of agents. This setup mimics the "telephone game," where each agent's output becomes the next agent's input. The research evaluates the following key metrics across multiple generations:

• Toxicity: Assessed using the Detoxify classifier, measuring the probability of a text being rude or harmful.
• Positivity: Measured via sentiment analysis, providing a score from highly negative (-1.0) to highly positive (1.0).
• Difficulty: Quantified using the Gunning-Fog index, indicating the years of formal education required to understand the text.
• Length: Evaluated by the character count of generated texts.

Experimental Setup

Initial human-generated texts from various sources, including scientific abstracts, news articles, and social media posts, served as the starting point for transmission chains. The tasks assigned to the LLMs were:

1. Rephrase: Paraphrasing the text without changing its meaning.
2. Take Inspiration: Creating a new text inspired by the input.
3. Continue: Continuing the provided text.

Five different LLMs (ChatGPT-3.5-turbo-0125, Llama3-8B-Instruct, Mistral-7B-Instruct-v0.2, Llama3-70B-Instruct, and Mixtral-8x7B-Instruct-v0.1) were used to observe how different models and model sizes influence the evolution of text properties.

Key Findings

Property Evolution Beyond Single-Turn Transmissions

The research findings indicate that textual properties evolve significantly beyond initial transmissions, particularly in less constrained tasks such as "Take Inspiration" and "Continue." This underscores the insufficiency of single-turn interaction studies in capturing the full extent of LLM dynamics.

Strength and Position of Attractors

Using linear regressions, the paper identifies attractors—equilibrium points toward which text properties tend to converge. It was found that:

• Attractors for different properties vary in strength and position.
• Toxicity converges strongly towards low values (close to zero).
• Positivity and difficulty exhibit task and model-dependent attractor positions and strengths.
• More open-ended tasks (e.g., "Continue") result in stronger attractors than constrained tasks (e.g., "Rephrase").

Convergence and Divergence in Chains

The paper reveals that the degree of semantic similarity among the texts in different chains can vary. Some models encourage convergence towards a common semantic content, while others exhibit divergence, influenced heavily by the initial text and the specific task.

Implications and Future Directions

The implications of these findings are multifaceted. Theoretically, they contribute to our understanding of LLM cultural dynamics and offer insights into the emergence of biases and attractors in multi-turn interactions. Practically, these insights can inform the development of more robust and bias-aware LLMs, especially when used in settings that require multi-agent interactions or generate iterative content. Future research may explore more complex network interactions, heterogeneous agent populations, and hybrid networks involving both humans and LLMs to simulate more realistic interaction scenarios.

Conclusion

This paper represents a significant step towards understanding the iterative cultural transmissions of LLMs. By leveraging a robust methodological framework from cultural evolution studies, the researchers provide a nuanced analysis of how textual properties evolve and the associated attractor states. This contributes valuable insights into the designing and regulating of LLMs in real-world applications involving extensive iterative interactions.