Chaotic Dynamics in Multi-LLM Deliberation

This paper models multi-LLM deliberation committees as random dynamical systems and demonstrates that even under deterministic temperature settings, factors like role differentiation and model heterogeneity induce chaotic instability characterized by positive empirical Lyapunov exponents, thereby establishing stability auditing as a critical requirement for reliable AI governance.

The Gist

Here is an explanation of the paper "Chaotic Dynamics in Multi-LLM Deliberation," translated into simple language with creative analogies.

The Big Idea: The "Unpredictable Committee"

Imagine you hire a committee of five AI experts to solve a difficult problem, like deciding how to fix a city's traffic or allocate a budget. You tell them, "Work together, argue your points, and vote." You run this meeting once, and they decide on Option A.

You think, "Great! Let's run the exact same meeting again with the exact same AI models and the exact same instructions." You expect them to come to the same conclusion, right?

This paper says: No, they probably won't.

Even if you set the "randomness" dial to zero (making the AI as deterministic as possible), the committee might decide on Option B the second time, and Option C the third time. The authors call this "Chaotic Dynamics." It means that tiny, invisible differences in how the AI processes information get amplified by the group discussion, leading to completely different outcomes every time you run the simulation.

---

The Two "Chaos Switches"

The researchers tested what makes these AI committees unstable. They found two main "switches" that turn the chaos on:

1. The "Specialized Roles" Switch (The Theater Analogy)

Imagine a group of friends hanging out. They are all just "friends." They talk, and they usually agree on where to eat dinner.
Now, imagine you force those same friends to act in a play. One must be the Director, one the Critic, one the Optimist, and one the Pessimist.

• The Finding: When the AI agents are given specific roles (like "Chair," "Welfare Expert," or "Security Expert"), the conversation becomes much more chaotic. The "Chair" tries to summarize, the "Security Expert" worries about risks, and the "Welfare Expert" looks at costs. These conflicting pressures amplify small disagreements into huge swings in opinion.
• The Metaphor: It's like a jazz band. If everyone just plays the same melody (no roles), it's stable. If everyone is told to play a specific, complex instrument with a specific solo (roles), the music can get wild and unpredictable.

2. The "Mixed Models" Switch (The Language Barrier Analogy)

Imagine a committee where everyone speaks the exact same dialect. They understand each other perfectly.
Now, imagine a committee where one person speaks English, one speaks French, one speaks a mix of both, and one speaks a dialect no one else has heard before.

• The Finding: When you mix different AI models (e.g., GPT, Claude, Gemini) in one committee, the chaos increases. Even if they are all trying to be helpful, they "think" and "phrase" things differently. These subtle differences in how they interpret the conversation cause the group to drift apart.
• The Metaphor: It's like a game of "Telephone" played by people who speak slightly different languages. The message gets distorted faster.

The Twist: The authors found that these two switches don't just add up; they interact in weird ways. Sometimes, having both roles and mixed models actually makes the system less chaotic than having just mixed models with no roles. It's a complex dance, not a simple math equation.

---

The "Chair" is the Wildcard

The researchers dug deeper to find who is causing the most trouble. They found that the Chair (the agent responsible for summarizing and guiding the conversation) is the main amplifier of chaos.

• The Analogy: Think of the Chair as the conductor of an orchestra. If the conductor is too active, trying to steer every note and summarize every solo, they might accidentally throw the whole orchestra off rhythm.
• The Fix: When the researchers removed the "Chair" role (letting the agents talk without a leader), the chaos dropped significantly. The group became more stable, even if they were still using mixed models.

---

The "Memory" Problem

Another finding was about how much history the AI remembers.

• The Setup: The AI agents were told to remember the last 15 minutes of conversation.
• The Fix: When the researchers told them to only remember the last 3 minutes (or even just 1 minute), the chaos decreased.
• The Metaphor: Imagine a group of people arguing. If they keep bringing up everything said 15 minutes ago, they get stuck in a loop of old arguments. If they only focus on what was just said, they move forward faster and settle on a decision more quickly.

---

Why Should You Care? (The "Governance" Warning)

This isn't just a cool science experiment; it's a warning for the future.

1. The "Deterministic" Myth: We often think that if we turn off the "randomness" setting (Temperature = 0) in AI, the results will be perfectly predictable. This paper proves that even with zero randomness, the system can still be unpredictable because of how the agents interact.
2. The Risk: If a hospital, a court, or a government uses an AI committee to make life-or-death decisions, they can't just run it once and trust the result. They might get a different answer if they run it again five minutes later.
3. The Solution: We need to "audit" the design of these AI committees. We need to check:
   • Are we giving them too many specific roles?
   • Are we mixing too many different AI models?
   • Are they remembering too much history?

The Takeaway

The paper concludes that stability is a design feature, not a default setting. If you want an AI committee that gives consistent answers, you have to carefully engineer how they talk to each other. You can't just throw five different AIs in a room and hope they agree. You have to design the room, the rules, and the memory so they don't accidentally drive the conversation into chaos.

Technical Summary

Here is a detailed technical summary of the paper "Chaotic Dynamics in Multi-LLM Deliberation" by Shimao, Khern-am-nuai, and Kim.

1. Problem Statement

The paper addresses a critical gap in the deployment of Collective AI Systems (multi-LLM committees). While these systems are increasingly used for governance and decision-making, their stability and reproducibility under repeated execution are poorly understood.

• The Core Issue: Even when running nominally identical committee configurations (same models, same prompts, same temperature $T=0$), the system can diverge into different trajectories and final decisions.
• The Risk: This unpredictability is not merely a "temperature artifact" (sampling noise) but appears to be structural. If institutions rely on these systems for policy, the inability to reproduce outcomes undermines governance, auditability, and trust.
• The Gap: Prior work has identified prompt-level instability or used "chaos" descriptively for failures. This paper aims to provide a dynamical systems framework to quantify, decompose, and map the specific architectural causes of this instability.

2. Methodology

The authors model the multi-LLM committee as a Random Dynamical System.

Experimental Design

• Setup: A 5-agent committee deliberates over 20 rounds on 12 distinct policy scenarios (covering immigration, health, climate, etc.).
• Protocol: A "Windowed-Summary" protocol where agents receive a sliding window of the last $k=15$ arguments and a state table. They output arguments and structured preference vectors $s^{(i)}_t = (p_A, p_B, p_C, \text{conf})$.
• Design Axes (2x2 Factorial):
  1. Role Structure: NoRoles (homogeneous agents) vs. Roles (agents assigned specific mandates: Chair, Welfare, Rights, Equity, Security).
  2. Model Composition: Uniform (all agents use the same model, e.g., GPT-4.1-mini) vs. Mixed (agents use different model families, e.g., GPT, Claude, Gemini, Grok).
• Conditions: Experiments are run primarily at Temperature $T=0$ to isolate structural instability from sampling noise.

Quantitative Metrics

• Trajectory Divergence ($D(t)$): The mean pairwise Euclidean distance between the committee's mean preference vectors across different replicate runs ($R=20$).
• Empirical Lyapunov Exponent ($\hat{\lambda}$): Calculated as the slope of $\log D(t)$ over rounds 3–20.
  • $\hat{\lambda} > 0$: Indicates exponential divergence (chaotic behavior).
  • $\hat{\lambda} \approx 0$: Indicates stable or convergent behavior.

Interventions & Ablations

• Role Ablation: Removing specific role mandates (e.g., the "Chair" role) to test their contribution to instability.
• Memory Depth Reduction: Reducing the context window size ($k$) from 15 to 3 or 1 to test if feedback loops drive the chaos.
• Semantic Perturbation: Testing if slight rephrasing of prompts (while preserving meaning) alters outcomes.

3. Key Contributions

1. Quantification of Instability: The first empirical demonstration that multi-LLM committees exhibit positive Lyapunov exponents ($\hat{\lambda} > 0$) even at $T=0$, proving instability is structural.
2. Decomposition of Instability Routes: Identification of two independent design routes that amplify divergence:
   • Route A (Institutional Differentiation): Assigning specific roles (especially the Chair) increases instability.
   • Route B (Compositional Heterogeneity): Mixing different model families increases instability.
3. Non-Additive Interaction: Discovery that combining both routes (Mixed Models + Roles) does not monotonically increase instability. In fact, the "Mixed+Roles" condition was less unstable than "Mixed+NoRoles," suggesting complex, non-additive interactions between design choices.
4. Mechanistic Insight: Identification of the Chair role as the dominant amplifier of chaos. Ablating the Chair role significantly reduced $\hat{\lambda}$.
5. Intervention Strategy: Demonstration that reducing memory depth (shortening the feedback loop) attenuates divergence, offering a path to stabilize these systems without collapsing deliberation into a single vote.

4. Key Results

Quantitative Findings (HL-01 Benchmark at $T=0$)

• Baseline (Uniform, NoRoles): $\hat{\lambda} = 0.0221$ (Positive, but low divergence).
• Route A (Uniform, Roles): $\hat{\lambda} = 0.0541$. Adding roles amplifies divergence.
• Route B (Mixed, NoRoles): $\hat{\lambda} = 0.0947$. Mixing models without roles causes high divergence.
• Interaction (Mixed, Roles): $\hat{\lambda} = 0.0519$. Surprisingly, adding roles to a mixed committee reduced the instability compared to the mixed/no-role condition, demonstrating non-additivity.

Mechanistic Findings

• The Chair Effect: In the "Roles" condition, the Chair agent is the primary source of instability.
  • Ablating the Chair role reduced $\hat{\lambda}$ most significantly (e.g., in IM-01, from 0.0838 to 0.0304).
  • Other roles (Welfare, Rights, etc.) had negligible or negative effects on $\hat{\lambda}$ when ablated.
  • The Chair's function of synthesizing diverse viewpoints and challenging arguments creates a feedback loop that amplifies minor initial perturbations.
• Memory Sensitivity: Reducing the memory window ($k$) from 15 to 3 or 1 consistently lowered $\hat{\lambda}$ across scenarios, confirming that feedback memory is a key amplifier of chaos.

Robustness

• Server-Side Non-Determinism: Even at $T=0$, cloud APIs exhibit floating-point non-determinism. The paper shows this infinitesimal noise is sufficient to trigger macroscopic divergence in these systems.
• Semantic Sensitivity: Slight rephrasing of prompts (synonym swaps, passive voice) also induced divergence ($\hat{\lambda} = 0.030$), though less than server-side noise.

5. Significance and Implications

• Governance & Auditability: The findings challenge the assumption that setting $T=0$ guarantees deterministic behavior. Institutions cannot rely on single-run outputs; stability auditing must become a core design requirement.
• Design Trade-offs: The paper warns against the naive assumption that "more diversity" (mixing models and roles) always leads to better or more stable outcomes. In fact, specific combinations can create highly unstable, unpredictable systems.
• Actionable Mitigation:
  • Protocol Design: Limiting memory depth (reducing the "history" agents can react to) can stabilize the system.
  • Role Management: The "Chair" role, while intended to facilitate consensus, acts as a chaos amplifier. Designers must weigh the benefits of synthesis against the risk of instability.
• Theoretical Contribution: It bridges social science (group polarization, framing effects) with dynamical systems theory (Lyapunov exponents, chaos), providing a rigorous mathematical framework for analyzing AI collective behavior.

Conclusion

The paper establishes that multi-LLM deliberation is inherently prone to chaotic dynamics driven by architectural choices. Instability is not an anomaly but a structural feature that can be induced by role differentiation and model heterogeneity. The authors provide a "stability map" showing that while these systems are sensitive to design, targeted interventions (like memory reduction or role ablation) can mitigate unpredictability, offering a path toward more robust AI governance systems.