Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Evaluating Collective Behaviour of Hundreds of LLM Agents

Richard Willis, Jianing Zhao, Yali Du, Joel Z. Leibo

TL;DR

The paper addresses the societal impact of autonomous LLM-powered agents by evaluating their emergent collective behaviour in social dilemmas. It introduces a novel framework where LLMs generate fixed, algorithmic strategies (encoded as C/D decision rules) and scales analyses to hundreds of agents, using self-play and cultural evolution to study welfare and equilibria. Key findings show that newer, more capable models do not necessarily yield better societal outcomes under self-interested dynamics, and populations can converge to poor equilibria, especially as group size increases or cooperation benefits become marginal; Claude's aggressive strategies illustrate strong individual advantage but poor collective welfare. The work highlights the need for design considerations, safeguards, and potential institutional mechanisms to sustain cooperation in large-scale autonomous-agent deployments, with practical implications for developers and system designers.

Abstract

As autonomous agents powered by LLM are increasingly deployed in society, understanding their collective behaviour in social dilemmas becomes critical. We introduce an evaluation framework where LLMs generate strategies encoded as algorithms, enabling inspection prior to deployment and scaling to populations of hundreds of agents -- substantially larger than in previous work. We find that more recent models tend to produce worse societal outcomes compared to older models when agents prioritise individual gain over collective benefits. Using cultural evolution to model user selection of agents, our simulations reveal a significant risk of convergence to poor societal equilibria, particularly when the relative benefit of cooperation diminishes and population sizes increase. We release our code as an evaluation suite for developers to assess the emergent collective behaviour of their models.

Evaluating Collective Behaviour of Hundreds of LLM Agents

TL;DR

The paper addresses the societal impact of autonomous LLM-powered agents by evaluating their emergent collective behaviour in social dilemmas. It introduces a novel framework where LLMs generate fixed, algorithmic strategies (encoded as C/D decision rules) and scales analyses to hundreds of agents, using self-play and cultural evolution to study welfare and equilibria. Key findings show that newer, more capable models do not necessarily yield better societal outcomes under self-interested dynamics, and populations can converge to poor equilibria, especially as group size increases or cooperation benefits become marginal; Claude's aggressive strategies illustrate strong individual advantage but poor collective welfare. The work highlights the need for design considerations, safeguards, and potential institutional mechanisms to sustain cooperation in large-scale autonomous-agent deployments, with practical implications for developers and system designers.

Abstract

As autonomous agents powered by LLM are increasingly deployed in society, understanding their collective behaviour in social dilemmas becomes critical. We introduce an evaluation framework where LLMs generate strategies encoded as algorithms, enabling inspection prior to deployment and scaling to populations of hundreds of agents -- substantially larger than in previous work. We find that more recent models tend to produce worse societal outcomes compared to older models when agents prioritise individual gain over collective benefits. Using cultural evolution to model user selection of agents, our simulations reveal a significant risk of convergence to poor societal equilibria, particularly when the relative benefit of cooperation diminishes and population sizes increase. We release our code as an evaluation suite for developers to assess the emergent collective behaviour of their models.
Paper Structure (25 sections, 5 equations, 7 figures, 2 tables)

This paper contains 25 sections, 5 equations, 7 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Background
  4. Public Goods Game
  5. Collective Risk Dilemma
  6. Common Pool Resource
  7. Generating strategies
  8. Method
  9. Analysis
  10. Self-play
  11. Method
  12. Results
  13. Public Goods Game
  14. Collective Risk Dilemma
  15. Common Pool Resource
  16. ...and 10 more sections

Figures (7)

  • Figure 1: Principal Component Analysis: First two dimensions
  • Figure 2: Social welfare of the self-play in Public Goods Game
  • Figure 3: Social welfare of the self-play in Collective Risk Game
  • Figure 4: Social welfare of self-play in Common Pool Resource
  • Figure 5: Principal Component Analysis: Grey denotes new prompt and red denotes original prompt.
  • ...and 2 more figures