Scale-Dependent Collective Adaptation in Self-Amending LLM Societies: A Cross-Family Study of Emergent Governance

arXiv:2605.17510 · nlin.AO, cs.MA · Submitted 2026-05-17 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Nonlinear Sciences. The day in nonlinear science: chaos, patterns, networks and adaptation.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "Scale-Dependent Collective Adaptation in Self-Amending LLM Societies".

Mira: Here's the gist of the paper. Scale-Dependent Collective Adaptation in Self-Amending LLM Societies: A Cross-Family Study of Emergent Governance is a study comparing how collective adaptation evolves across different model sizes and families when the rules of play are subject to collective amendment, revealing a non-monotonic relationship between model scale and institutional change The paper finds that collective adaptation does not improve monotonically with model size, but instead, both LLM families exhibit a narrow mid-scale regime that supports sustained rule adoption and balanced consensus Collective Adaptation Across Scales The study investigates group decision-making in artificial societies where the rules of play are themselves subject to collective amendment using the self-amending game Nomic

Title and authors: Kai: Let's look at what they actually did in this paper on "Scale-Dependent Collective Adaptation in Self-Amending LLM Societies: A Cross-Family Study of Emergent Governance." The focus is on the self-amending game Nomic to see how these LLM societies evolve their own rules.

Mira: They are using a controlled testbed, varying model scale within Qwen3 point 5 and Gemma3 families—ranging from two hundred seventy million up to twenty-seven billion parameters—while keeping the underlying architecture and training methods fixed for each family.

Lev: So they held the core structure constant but varied the size of the agent population, which is a very specific way to isolate scale effects from architectural differences.

Kai: That’s right. They pair this behavioral measurement with analyzing hidden states using something called hidden-state probing to see how the internal representation changes as rules are being amended.

Mira: It’s not about attributing genuine understanding there, but measuring patterns like chain-of-thought prompting or theory of mind that develop unevenly across those scales.

Lev: That means we’re looking at functional metrics—how well the system performs a task—rather than whether the AI actually *understands* why it changed its rule.

Kai: And the finding they drive home is that collective adaptation doesn't improve monotonically with model size, which is counter-intuitive if you just think bigger models always mean better coordination.

Mira: The paper emphasizes that the non-monotonicity itself—the U-shaped or inverse scaling regimes—is what’s important, not just finding one single scale where things look perfect.

Lev: So the implication for us in building systems is that we should be designing governance mechanisms to account for this dip and rise, not just optimizing for the biggest possible parameter count.

The paper's summary: Kai: To summarize what they found regarding collective adaptation across these LLM societies, the main takeaway is that there’s a non-monotonic relationship between model size and institutional change.

Mira: Specifically, both the Qwen3 point 5 and Gemma3 families exhibit a narrow mid-scale regime that supports sustained rule adoption, diverse amendments, and balanced consensus for governance <ref:2605.17510#pg1,families exhibit a narrow mid-scale regime that supports sustained rule adoption>.

Lev: That mid-scale optimum acts like a stable operational sweet spot where the coupling between strategies and institutions works best before things get too rigid or too chaotic at the extremes.

Kai: Smaller models tend to remain rule-inert, meaning they don't change their rules much, while larger models often converge on a more restricted set of rules.

Mira: The crucial part is that this pattern—the non-monotonic shape itself—is what’s being studied across the two distinct model families, which separates shared regularities from stuff specific to one architecture.

Lev: So we're not just looking at one system; we're seeing a structural property of how these agents interact when the environment they are operating in is mutable.

Kai: It shows that the way an AI society evolves its own governing rules has a specific scaling behavior that depends on where you sit on the scale, not just being at the very top or bottom.

Mira: This moves the discussion away from asking "which model is best" to asking "what scale supports stable governance?"

The paper's improvements: Kai: The authors actually suggest several ways we should look at this research and how we apply it to things like designing institutional tasks. First, they argue that model selection for institutional tasks should be scale dependent, not just based on raw model size.

Mira: They propose using that mid-scale regime to support sustained rule adoption and ensure diverse amendments happen while maintaining a balanced consensus among agents.

Lev: That means when we design an environment for AI to govern itself, we need to build in mechanisms that encourage movement through that sweet spot, not just aiming for maximum complexity.

Kai: Second, they suggest implementing a way to detect the specific operational regime of an LLM society—whether it’s frozen, explorative, or oscillatory—by looking at the shapes of the rule trajectory itself.

Mira: That lets us tailor our interventions based on whether we want convergence on a working institution or if we anticipate chronic revisitation of old settings.

Lev: So instead of just checking a final state, we need to analyze the path the rules took to get there—is it settling down, or is it bouncing around?

Kai: Third, they propose developing vote-prediction probing techniques that target intermediate-to-mid hidden layers to find that specific "behavioural-coupling pattern."

Mira: This pattern links linearly decodable vote outcomes with balanced downstream voting, and the paper notes this only happens at those specific sweet spots.

Lev: That’s interesting because it suggests we can actually probe the internal workings of the AI society to pinpoint exactly where that coupling exists before we try to let it self-govern.

Kai: And they suggest using hidden state divergence profiles for a dissociation analysis, specifically looking for a "mid-layer plateau" at that sweet spot instead of seeing just a late-layer surge when we perturb the institution.

Conclusion: Mira: So wrapping up on this paper, the implications are that we need to treat model scale not as a single variable, but as something context-dependent for institutional evolution in AI societies. The core finding is that collective adaptation isn't smooth across all sizes; it has these distinct regimes.

Lev: What this means for real hardware running these simulations is that our error correction overhead needs to be tuned to handle the dynamics of a system operating near that mid-scale sweet spot, not just scaling up blindly.

Kai: Exactly. We need mechanisms to detect those specific operational regimes and perhaps even probe those intermediate layers because that’s where the interesting coordination happens.

Mira: The paper warns us about heterogeneity as well; it says heterogeneous societies in both families collapse into asymmetric distrust, where vetoes from the largest member override any member at the sweet spot.

Lev: That’s a serious caveat for governance design—we have to build in ways to mitigate that veto bias from dominant agents if we want stable collective adaptation.

Kai: Overall, this study on "Scale-Dependent Collective Adaptation in Self-Amending LLM Societies: A Cross-Family Study of Emergent Governance" tells us that the path to AI self-governance isn't a simple scaling problem; it’s about finding the right operational regime at the right scale.

Mira: It shifts our focus from just optimizing model size to understanding the dynamics of rule rewriting and consensus building in these artificial systems.

Lev: We need more work on how to formally model that mid-scale plateau you mentioned, because if we can map it, we might be able to build more robust agent-based governance structures.

Kai: Right. Next time, I think we should look at how these concepts apply when the rules themselves are being rewritten by something other than just simple amendments within the game Nomic.

Center for Brain Science, RIKEN · Department of Evolutionary Biology and Environmental Studies, University of Zurich · Information R&D and Strategy Headquarters Information R&D and Strategy Headquarters · School of Knowledge Science, Japan Advanced Institute of Science and Technology · Valencian Research Institute of Artificial Intelligence Universitat Politecnica de Valáncia · The Psychometrics Centre, University of Cambridge

nlin.AO, cs.MA

Submitted: 2026-05-17

Updated: 2026-11-07

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 91/100

The gist: Scale-Dependent Collective Adaptation in Self-Amending LLM Societies: A Cross-Family Study of Emergent Governance is a study comparing how collective adaptation evolves across different model sizes…

Key concepts

Collective Adaptation
This refers to how a group of LLMs collectively changes their rules of play over time in response to new information or interactions. The research examines whether this process gets better or worse as the group gets larger.
Self-Amending Game Nomic
This is the specific game framework used in the study, designed to model artificial societies where the participants (LLMs) can collectively propose and adopt changes to their own governing rules. It tests how these rule changes affect system stability.
Non-monotonic Relationship
This means that as one factor—in this case, model size—increases, the outcome does not consistently go up or down. Instead, the effect is complex: adaptation might improve up to a certain point before potentially declining or stabilizing at a specific optimal level.
Mid-scale Regime
This is the specific range of model sizes where collective adaptation performs best. The research suggests that this intermediate size supports stable rule adoption and consensus, contrasting with the findings that very small or very large models might lead to different governance outcomes.

Terminology

Summary

Scale-Dependent Collective Adaptation in Self-Amending LLM Societies: A Cross-Family Study of Emergent Governance is a study comparing how collective adaptation evolves across different model sizes and families when the rules of play are subject to collective amendment, revealing a non-monotonic relationship between model scale and institutional change

The paper finds that collective adaptation does not improve monotonically with model size, but instead, both LLM families exhibit a narrow mid-scale regime that supports sustained rule adoption and balanced consensus

Collective Adaptation Across Scales

The study investigates group decision-making in artificial societies where the rules of play are themselves subject to collective amendment using the self-amending game Nomic

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Model selection for institutional tasks should be scale dependent rather than model size dependent, as collective adaptation does not improve monotonically with model size. This suggests using a mid-scale regime to support sustained rule adoption, diverse amendments, and balanced consensus.

  2. Implement a mechanism to detect the specific operational regime of an LLM society (frozen, explorative, or oscillatory) by analyzing rule trajectory shapes rather than just summary statistics. This allows for tailored interventions based on whether the goal is convergence at a working institution or chronic revisitation of prior settings.

  3. Develop vote-prediction probing techniques targeting intermediate-to-mid hidden layers to identify the behavioural-coupling pattern that links linearly decodable vote outcomes with balanced downstream voting, which occurs only at specific sweet spots.

  4. Use hidden state divergence profiles as a dissociation analysis to rule out simple magnitude interpretations of internal responses; specifically, look for a mid-layer plateau at the sweet spot instead of a late-layer surge when analyzing institutional perturbations.

  5. Design governance systems that account for heterogeneous society collapse by recognizing that heterogeneous societies in both families collapse into asymmetric distrust where vetoes from the largest member override any sweet-spot member that is present. This requires mechanisms to mitigate veto bias from dominant agents.

Abstract

We study group decision-making in artificial societies where the rules of play are themselves subject to collective amendment. Using the self-amending game Nomic, we compare multiple scales across two LLM families and find that collective adaptation does not improve monotonically with model size. Instead, both families exhibit a narrow mid-scale regime that supports sustained rule adoption, diverse amendments, and balanced consensus. Smaller models tend to remain rule-inert, whereas larger models often converge on restrictive voting patterns, and heterogeneous mixed-size groups collapse into veto-driven gridlock. These cross-scale contrasts persist under temperature perturbations and under a shift from unanimity to majority voting, although latent-state structure varies by family and scale. Hidden-state divergence alone does not explain collective performance: high representational divergence can coincide with poor behavioural outcomes. Linear probes reveal regime-selective coupling between latent vote-predictive signals and collective behaviour, but decodability is necessary rather than sufficient for adaptive play. Overall, the recurring regularity is non-monotonicity, not the particular scale at which the optimum appears. Self-amending games therefore provide a controlled testbed for studying collective adaptation in artificial societies beyond raw model scale.

Sources

Related papers