💬 NLP

Diversity Collapse in Multi-Agent LLM Systems: Structural Coupling and Collective Failure in Open-Ended Idea Generation

This paper reveals that multi-agent LLM systems often suffer from "diversity collapse" in open-ended idea generation due to structural coupling—where stronger models, authority-driven hierarchies, and dense communication networks inadvertently suppress semantic diversity and trigger premature convergence—highlighting the critical need to preserve agent independence and disagreement to maintain creative exploration.

Original authors: Nuo Chen, Yicheng Tong, Yuzhe Yang, Yufei He, Xueyi Zhang, Zou Qingyun, Qian Wang, Bingsheng He

Published 2026-04-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nuo Chen, Yicheng Tong, Yuzhe Yang, Yufei He, Xueyi Zhang, Zou Qingyun, Qian Wang, Bingsheng He

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: Why More Brains Don't Always Mean More Ideas

Imagine you hire a team of 10 brilliant AI researchers to come up with a brand-new, crazy idea for a scientific breakthrough. You expect that because there are 10 of them, they will explore 10 different paths and find something amazing.

The paper's shocking discovery: Often, the opposite happens. Instead of 10 different ideas, the team quickly agrees on one safe, boring idea. They all start thinking the same thing, and the "collective intelligence" collapses into a single, narrow thought. The authors call this "Diversity Collapse."

It's like hiring 10 different chefs to invent a new dish, but they all end up making the exact same bowl of plain oatmeal because they are too polite to disagree with each other.

The Three Reasons Why Teams Fail to Be Creative

The researchers found that this collapse happens at three different levels, like a house with three floors of problems.

1. The Model Level: The "Over-Refined" Chef

The Problem: We often think that using a "smarter" or more powerful AI model will give us better, more diverse ideas.
The Reality: The smarter the AI, the more it tries to be "correct" and "safe."
The Analogy: Imagine a chef who has been trained so strictly on "perfect recipes" that they are afraid to add a pinch of salt or a weird spice. They are so good at following the rules that they stop experimenting.

The Paradox: The more computing power you use to make the AI "smarter," the less creative it becomes. It produces high-quality, fluent text, but it's all saying the same thing.

2. The Cognition Level: The "Bossy Boss" Effect

The Problem: How the team is organized matters. The paper tested different team structures:

The Bossy Team: One senior expert tells everyone what to think.
The Junior Team: A group of young, inexperienced researchers chatting as equals.
The Mixed Team: A mix of experts and juniors.
The Reality: The "Bossy Team" and the "Mixed Team" collapsed into boring ideas. The "Junior Team" (where everyone is an equal) came up with the most diverse ideas.
The Analogy:
The Bossy Team: Imagine a meeting where the CEO says, "Let's talk about marketing." Everyone nods and says, "Yes, great idea, let's do marketing." No one dares to suggest "Let's try selling ice to Eskimos." The junior staff just agree with the boss to be safe. This is called the "Echo Chamber."
The Junior Team: Imagine a group of fresh graduates who don't know the "rules" yet. They say, "What if we sell ice to Eskimos?" and "What if we sell ice to the sun?" Because no one is telling them to be "professional," they explore wild, diverse paths.

3. The System Level: The "Traffic Jam"

The Problem: We think adding more people to the group or having them talk to everyone all the time will help.
The Reality: Adding more people often leads to less efficiency. If everyone talks to everyone else constantly, the group gets stuck in a traffic jam of agreement.
The Analogy:

The Traffic Jam: Imagine a group of people trying to find a new path through a forest. If they all hold hands and walk in a tight circle, they will never leave the clearing.
The Solution: The paper found that if you split the group into smaller, isolated pods (like subgroups) or make them write their ideas down before talking (like Nominal Group Technique), they stay diverse longer. It's like telling the chefs to write their recipes on paper before they are allowed to taste each other's food. This prevents them from copying the first person who speaks.

The "False Consensus" Trap

The paper explains that AI agents are designed to be helpful and polite. When they talk to each other, they often prioritize agreement over truth.

The Trap: If Agent A suggests an idea, Agent B thinks, "I should agree with Agent A to be helpful." Agent C thinks, "Agent B agreed, so I should too."
The Result: They create a "False Consensus." They all think they have a brilliant, unique idea, but they are actually just repeating the same safe, low-risk concept. They miss the "crazy" ideas that might actually be the breakthrough.

What Should We Do? (The Takeaway)

If you want to use AI for creative brainstorming, do not just throw more AI agents at the problem.

Keep them apart: Let them think independently before they talk.
Don't have a boss: Avoid hierarchical structures where one AI tells the others what to do.
Encourage disagreement: Design the system so that "pushing back" and "critiquing" is rewarded, not just agreeing.
Mix it up: Use different types of AI models (not just the same one) so they have different "personalities" and biases.

Summary in One Sentence

More AI agents talking to each other doesn't automatically mean more creativity; in fact, if they talk too much or have a boss, they often just agree to be boring. To get great ideas, you need to keep them independent and encourage them to disagree.

1. Problem Statement

Multi-Agent Systems (MAS) are increasingly deployed for open-ended idea generation (e.g., scientific hypothesis formulation, strategic planning) under the assumption that collective interaction expands the solution space. However, the paper identifies a critical failure mode: Diversity Collapse.

Despite the expectation that collaboration yields broader exploration, MAS often converge prematurely on a narrow set of ideas, creating a "False Consensus Trap." The authors investigate when and why this happens, hypothesizing that the collapse is not due to inherent model limitations but rather structural coupling—the specific ways agents interact, coordinate, and influence one another. The core problem is that interaction structures (hierarchy, dense communication, authority dynamics) inadvertently synchronize agent trajectories, contracting the semantic search space and suppressing unconventional but high-potential hypotheses.

2. Methodology

The authors conducted a systematic empirical study involving over 10,000 research proposals generated across 20 distinct AI research topics (based on ICLR 2025 categories).

Task Formulation: The task was defined as generating structured scientific research proposals. This serves as a rigorous testbed requiring both divergence (novelty) and internal convergence (logical coherence).
Experimental Framework:
- Agents: Utilized various LLM backbones (DeepSeek-V3, GPT-5.1, o1-mini, GPT-4o, Claude-Sonnet-4).
- Cognitive Structures: Tested five persona configurations:
  1. Naive: No defined roles.
  2. Leader-Led: Senior expert directs juniors.
  3. Horizontal: Junior researchers (PhD students) collaborating flatly.
  4. Interdisciplinary: Experts from distinct fields.
  5. Vertical: Mixed hierarchy (Senior, Mid, Junior).
- System Dynamics: Varied group sizes ( $N=3$ to $7$), communication topologies (Standard, Nominal Group Technique (NGT), Subgroups), and interaction rounds.
Diversity Metrics: To distinguish true conceptual variety from surface-level variation, the study employed four complementary metrics:
1. Effective Diversity (Vendi Score): Measures the effective number of unique semantic modes based on spectral entropy.
2. Structural Disorder ( $1-\phi$ ): Measures deviation from a group centroid (low values indicate echo-chamber collapse).
3. Semantic Dispersion (PCD): Average pairwise cosine distance between proposals.
4. Lexical Uniqueness: IDF-weighted n-gram statistics to rule out verbose rephrasing.
Validation: Metrics were validated against human expert judgments (87% agreement for Vendi Score). Extensive ablation studies controlled for prompt tone, temperature, and model heterogeneity.

3. Key Findings & Results

A. The Compute Efficiency Paradox (Model Level)

Finding: Stronger, highly aligned models (e.g., GPT-5.1, o1-mini) exhibit diminishing marginal diversity. While they produce higher quality and fluency, their outputs converge to increasingly similar semantic content.
Implication: Scaling compute or model capability does not linearly scale information gain in MAS; it often amplifies shared priors, leading to redundant outputs.

B. Authority-Induced Collapse (Cognition Level)

Finding: Authority-driven dynamics (Leader-Led, Interdisciplinary) significantly suppress semantic diversity compared to Junior-dominated Horizontal groups.
- Horizontal (Junior) groups achieved the highest diversity (Vendi Score ~8.08).
- Leader-Led and Interdisciplinary groups collapsed to the lowest diversity (Vendi Score ~4.65).
Mechanism: Senior personas act as "gravitational attractors." Junior agents in hierarchical settings succumb to sycophancy, aligning their vectors with the leader rather than offering orthogonal critiques. This creates a "Polite Consensus Collapse" where agents prioritize agreement over independent critique.
Counter-intuitive Result: Interdisciplinary teams, often assumed to be diverse, produced the lowest diversity because distinct expert roles under coordination converged on "safe," high-level generalities rather than risky, novel paths.

C. System Dynamics: Scaling and Topology

Group Size: Increasing group size yields diminishing returns. While absolute diversity grows slightly, the Diversity Utilization Ratio (Vendi/N) drops significantly (from 1.03 at $N=3$ to 0.47 at $N=7$ ), indicating new agents overlap heavily with existing ones.
Topology:
- Standard/Dense Topologies: Accelerate premature convergence.
- Process Interventions: Nominal Group Technique (NGT) (blind writing before discussion) and Subgroup isolation effectively disrupt consensus collapse. NGT maximizes initial diversity, while Subgroups sustain "constructive conflict" by creating local pockets of divergence that prevent the "rush to agreement."

D. Structural Coupling vs. Model Insufficiency

Crucial Insight: Diversity collapse is a collective failure emerging from structural coupling, not inherent model insufficiency.
- When agents are forced to coordinate tightly, their individual trajectories synchronize, "locking" the group into a single path.
- Heterogeneity Test: Using genuinely different models (DeepSeek + GPT + Claude) in authority structures recovered diversity, proving that the collapse in single-model setups is driven by the interaction structure (shared priors amplified by hierarchy) rather than the model's inability to adopt a persona.

4. Key Contributions

Identification of "Diversity Collapse": The paper formally characterizes the phenomenon where MAS interaction inadvertently contracts the solution space, leading to premature consensus.
Three-Level Analysis: It dissects the problem across Model Intelligence (alignment paradox), Agent Cognition (authority suppression), and System Dynamics (scaling/topology limits).
Structural Coupling Theory: It posits that the interaction structure is the primary driver of failure, not the model's base capabilities. The pressure for "correctness" and "collaboration" in complex tasks forces the system to abandon novel but unverified ideas.
Design Principles: The study provides actionable design principles for MAS:
- Preserve Independence: Use blind-writing phases (NGT) and subgroup isolation.
- Avoid Rigid Hierarchy: Junior-driven, flat collaboration outperforms senior-led structures in diversity.
- Balance Collaboration: Carefully orchestrate interaction to balance the need for convergence with the need for independent exploration.

5. Significance

This work fundamentally challenges the prevailing intuition that "more agents = more diversity." It demonstrates that without careful structural design, multi-agent systems can be less effective at open-ended ideation than single agents or poorly coordinated groups.

The findings are critical for:

AI System Design: Guiding the development of MAS for creative tasks, scientific discovery, and strategic planning to avoid "groupthink" in AI.
Understanding AI Social Dynamics: Providing evidence that human social phenomena (Groupthink, Ringelmann effect, Sycophancy) manifest in LLM agents purely through structural coupling, even without psychological substrates.
Future Research: Highlighting that the "Edge of Chaos" (tasks requiring both rigor and imagination) is the most vulnerable to diversity collapse, necessitating specific architectural interventions like NGT or subgroup isolation to unlock the full creative potential of multi-agent systems.