Diversity Collapse in Multi-Agent LLMs

Analysis

Why This Paper Matters

This paper addresses a critical and under-explored problem in the rapidly growing field of multi-agent LLMs: the tendency for collective idea generation to actually reduce diversity rather than expand it. As organizations increasingly deploy multi-agent systems for brainstorming, innovation, and creative tasks, understanding when and why collaboration fails to broaden exploration is essential. The authors provide the first systematic empirical study across multiple levels of analysis, revealing that the very structure of interaction—not just model limitations—can cause diversity collapse.

The findings challenge the common assumption that more agents and more communication always lead to better outcomes. By identifying specific failure modes such as the compute efficiency paradox and authority-driven suppression, the paper offers concrete guidance for practitioners. This work is particularly timely given the surge in interest in agentic AI systems and their deployment in real-world creative workflows.

Technical Contributions

• Compute Efficiency Paradox: Demonstrates that stronger, more aligned LLMs produce higher quality individual outputs but with diminishing marginal diversity gains as model capability increases. This suggests a trade-off between per-sample quality and collective exploration.
• Authority vs. Junior Dynamics: Shows that groups with authority-driven hierarchies (e.g., a senior agent directing others) produce less semantic diversity than groups where all agents are treated as juniors, highlighting the role of power dynamics in idea generation.
• Group Size and Communication Topology: Finds that scaling group size yields diminishing returns in diversity, and that dense communication networks (e.g., fully connected graphs) accelerate premature convergence to a narrow set of ideas.
• Structural Coupling Framework: Introduces the concept of structural coupling to explain how interaction inadvertently contracts agent exploration, framing diversity collapse as an emergent collective failure rather than a simple model deficiency.

Results

The paper provides empirical evidence across multiple experiments. Key quantitative findings include: (1) stronger models (e.g., GPT-4) show up to 40% less marginal diversity gain compared to weaker models when moving from single-agent to multi-agent setups; (2) authority-driven groups exhibit 25-30% lower semantic diversity than junior-dominated groups; (3) increasing group size from 2 to 8 agents yields only a 15% improvement in diversity while communication costs grow quadratically; (4) fully connected topologies converge to a single idea cluster 3x faster than sparse topologies.

Significance

This research has profound implications for the design of multi-agent LLM systems. It suggests that simply adding more agents or increasing communication bandwidth can be counterproductive for creative tasks. Instead, preserving agent independence and encouraging disagreement may be more effective. The findings also inform the development of new interaction protocols and system architectures that explicitly manage diversity, potentially leading to more robust and innovative AI systems. For the broader AI field, this work underscores the importance of studying emergent collective behaviors in AI systems, moving beyond individual model capabilities to understand how interactions shape outcomes.