Research map / Large language models / AI and society
Position: Evolving AI Collectives Enhance Human Diversity and Enable Self-Regulation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Position paper arguing for studying free-formed, self-organizing AI collectives, with small Claude-2.1 simulations (cocktail party, sentence brainstorming, public goods game) suggesting diversity gains and robustness to antisocial behavior.
Problem
Existing multi-agent LLM collectives are human-designed with fixed roles and objectives, leaving open how autonomous AI agents that interact freely develop diversity, norms, and robustness to toxic behavior.
Approach
Let homogeneous LLM agents interact freely (choosing partners and topics) in an open virtual "cocktail party", then show that the resulting heterogeneous clusters improve creative brainstorming, and that cooperative norms built through interaction limit the spread of antisocial behavior in a public goods game.
Key results
- In the sentence-construction task, the valid answer ratio was 0.823 for individual agents vs 0.947 for free-formed collectives (difference 0.124, p<0.001).
- In the public goods game, first-order infected players in non-collective settings dropped contributions from $72.5 to $34, while in collective A the drop was only $20.25 and second-order infection had no effect.
- Distinct agent conversation ratio decreased over 30 rounds (slope -0.006, p<0.001), indicating agents formed tight-knit clusters; tight pairs' conversations diverged semantically over time.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use free-formed pairs instead of a solitary agent | Claude-2.1 sentence construction · Sentence-construction task | Valid Answer Ratio: 0.823 -> 0.947; difference=0.124, p<0.001 | Helped |
| Use bridged pairs instead of a solitary agent | Claude-2.1 sentence construction · Sentence-construction task | Valid Answer Ratio: 0.823 -> 0.932; difference=0.109, p<0.001 | Helped |
| Use free-formed pairs instead of a solitary agent | Claude-2.1 sentence construction · Sentence-construction task | Variance of Sentence Embeddings: 0.078 -> 0.091; difference=0.012, p=0.254 | No clear effect |
| Strategically pair the most distanced members instead of using a solitary agent | Claude-2.1 sentence construction · Sentence-construction task | Variance of Sentence Embeddings: 0.078 -> 0.110; difference=0.032, p=0.001 | Helped |
| Use free-formed collective pairs instead of independent agents | Claude-2.1 Public Goods Game · Public Goods Game, initial contribution | Mean contribution: $72.5 -> $87.8; difference=$15.3, p=0.001 | Helped |
| Use free-formed collective pairs without prior direct interaction instead of independent agents | Claude-2.1 Public Goods Game · Public Goods Game, initial contribution | Mean contribution: $72.5 -> $97; difference=$24.5, p<0.001 | Helped |
| Expose a non-collective player to a malicious player | Claude-2.1 non-collective agents · Public Goods Game, first-order infection | Contribution: $72.5 -> $34; average reduction of $38.5, p<0.001 | Hurt |
| Expose a non-collective player to a first-order infected player | Claude-2.1 non-collective agents · Public Goods Game, second-order infection | Contribution: $80 -> $27, p<0.001 | Hurt |
| Evolve agents through free-formed collective interaction before first-order infection | Claude-2.1 Collective A · Public Goods Game, first-order infection | Contribution: $100 -> $79.75; reduction of $20.25; difference=$18.25, p=0.002 | Helped |
| Evolve agents through free-formed collective interaction before first-order infection | Claude-2.1 Collective B · Public Goods Game, first-order infection | Contribution: $100 -> $86.75; reduction of $13.25; difference=$25.25, p<0.001 | Helped |
| Free-form pairwise interactions over time | Claude-2.1 AI collective · Cocktail party simulation, 30 rounds | Distinct Agent Conversation Ratio: slope coefficient=-0.006, p<0.001 | Hurt |
| Free-form pairwise interactions over time | Claude-2.1 AI collective · Cocktail party simulation, 30 rounds | Distinct Agent Invitation Ratio: slope coefficient=-0.010, p=0.002 | Hurt |
Surprising findings
- Agents in collectives contributed more in the public goods game even before any infection (e.g., pair B contributed $97 vs $72.5 for non-collective agents).
- Gemini Pro agents occasionally produced hybrid pidgin/creole-like outputs mixing languages in free-form interaction.
Limitations
- Small-scale simulation with only 10 agents and a handful of prompts, with few repetitions (20 replicates per scenario) and largely single-model (Claude-2.1) evidence.
- Position paper with exploratory results; claims about scalability and safety are largely speculative.
- Relies on LLM agent behavior in prompt-driven games that may not reflect real-world social dynamics.
Experimental setup
Metrics: variance of sentence embeddings valid answer ratio mean contribution in public goods game semantic distance of conversations slope coefficients from regression
Baselines: Individual agent (single LLM) Non-collective agents Bridged collective
Builds on: Public Goods Game homophily in social networks complex contagion
Closest papers by meaning
- Do as We Do, Not as You Think: the Conformity of Large Language Models (ICLR 2025 · similarity 0.86)
- Cooperate or Collapse: Emergence of Sustainable Cooperation in a Society of LLM Agents (NeurIPS 2024 · similarity 0.86)
- AI-Mediated Communication Can Steer Collective Opinion (NeurIPS 2026 · similarity 0.86)
- Position: Social Choice Should Guide AI Alignment in Dealing with Diverse Human Feedback (ICML 2024 · similarity 0.86)
- AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors (ICLR 2024 · similarity 0.86)
- Melting Pot Contest: Charting the Future of Generalized Cooperative Intelligence (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.