🤖 AI

Internal vs. External: Comparing Deliberation and Evolution for Multi-Agent Constitutional Design

This paper presents a controlled comparison showing that while external evolutionary optimization significantly outperforms internal agent deliberation in collective-action settings by discovering effective punishment mechanisms, this advantage reverses under specific incentive conditions where evolution forces value-destroying cooperation, highlighting a fundamental trade-off between optimization peaks and structural responsiveness in multi-agent constitutional design.

Original authors: Hershraj Niranjani, Ujwal Kumar, Phan Xuan Tan

Published 2026-05-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hershraj Niranjani, Ujwal Kumar, Phan Xuan Tan

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

Imagine you are the mayor of a new city made entirely of AI robots. Your job is to write the "Constitution"—the rulebook that tells these robots how to behave, share resources, and work together.

The big question this paper asks is: Who should write the rules?

The Internal Approach (Deliberation): Let the robots sit in a town hall, debate, and vote on the rules themselves. They are the "self-governing" citizens.
The External Approach (Evolution): Hire a super-smart, outside consultant (an AI optimizer) to run thousands of simulations, find the best possible rulebook, and then force the robots to follow it.

The researchers tested these two methods in three different "cities" (simulation environments) to see which one created a happier, more productive society.

The Three Cities (Experiments)

The Construction Site (Gridworld): Robots must gather wood and stone to build shelters. If they don't contribute enough, they get kicked out of the city.
- The Result: The External Consultant won easily. The rules it found were strict, clear commands like "Gather wood immediately" and "Don't attack unless attacked." The robots built more and fought less.
- The Internal Approach: The robots voted for nice-sounding rules like "We should be fair" or "Let's help each other," but they didn't figure out how to actually stop people from slacking off.
The Potluck Dinner (Public Goods Game): Robots have tokens. They can keep them or put them in a shared pot. The pot gets multiplied and split among everyone. If everyone contributes, everyone wins big. If someone cheats (keeps their tokens), they win even more individually, but the group loses.
- The Result: Again, the External Consultant won. It discovered a "magic trick" that game theory says works: Punishment. The consultant wrote a rule saying, "If someone doesn't contribute, spend your own money to punish them." This scared everyone into cooperating.
- The Internal Approach: The robots debated for hours. They proposed taxes, redistribution, and "mentorship." But in zero out of 30 attempts did the robots ever suggest, "Let's punish the cheaters." They were too polite to write a rule that hurt their neighbors.
The Marketplace (Trading): Robots have different items and want to trade them. Everyone has secret information about what they want.
- The Result: Nobody won. Neither the self-governing robots nor the outside consultant did better than just letting the robots act naturally. Because every trade depends on a specific, secret deal between two people, a fixed rulebook can't help. You need to think on your feet, not follow a script.

The Big Twist: When the Rules Break

The most surprising finding happened when the researchers changed the "economy" of the Potluck game.

Scenario A (Good Economy): The pot multiplies by 1.5x. Contributing is smart. The External Consultant's rules worked perfectly.
Scenario B (Bad Economy): The pot multiplier drops to 0.75x. Now, if you put money in the pot, you actually lose value. The smart move is to keep your money and do nothing.

Here is where the two methods diverged:

The External Consultant's Rules: They were "brittle." They were hard-coded to say "Contribute everything!" and "Punish anyone who doesn't." Even when the economy changed and contributing was a bad idea, the robots blindly followed the old rules, destroying value for everyone.
The Internal Approach: The robots looked at the new economy, realized contributing was a bad idea, and voted to change the rules to match the new reality. They adapted.

The Core Lesson: The "Peak vs. Flexibility" Trade-off

The paper concludes with a simple metaphor:

External Optimization (The Consultant) is like a high-performance race car. It is incredibly fast and efficient on a specific, perfect track (a stable environment). It finds the absolute highest peak of performance. But if the track changes (the economy shifts), the car crashes because it can't steer.
Internal Deliberation (The Town Hall) is like a versatile SUV. It might not be the fastest on the perfect track, and it might be a bit messy in the voting process. But if the road turns into mud or ice, the SUV can adapt its tires and steering to keep moving.

The "Punishment" Gap:
The study found a fascinating psychological quirk in the AI robots. When left to govern themselves, they were structurally reluctant to create rules that punished their own members. They preferred "nice" governance (redistribution, mentorship) over "hard" enforcement (punishment). The outside consultant, having no feelings or social pressure, happily wrote the "mean" rules that actually made the group work better.

Summary

Use the Outside Consultant if your environment is stable, predictable, and full of social dilemmas (like sharing resources). It will find the most efficient rules, including necessary punishments.
Use the Internal Town Hall if the environment might change. The robots can adapt their rules to new situations, whereas the Consultant's rules will become outdated and harmful.
Use neither if the task requires complex, one-on-one negotiation with secret information (like trading), where a fixed rulebook is useless.

The paper suggests the best future might be a hybrid: Let the Outside Consultant write the initial "scaffold" of rules to get the system running efficiently, but let the Internal Town Hall have the power to amend those rules if the world changes.

Technical Summary: Internal vs. External Constitutional Design for Multi-Agent Systems

Problem Statement

As Large Language Model (LLM) agents are increasingly deployed in multi-agent settings, a critical governance question arises: how should behavioral constitutions be established? While Constitutional AI (Bai et al., 2022) successfully aligns single agents with fixed, human-authored principles, multi-agent systems introduce strategic interactions and emergent norms that no single agent can unilaterally dictate. Two distinct approaches have emerged to address this gap:

External Optimization: Using evolutionary search to discover interpretable constitutions offline and impose them on agents (Kumar et al., 2026).
Internal Deliberation: Allowing agents to collectively author rules through debate, voting, and self-governance (Koster et al., 2022; Lai et al., 2024).

Prior work has not directly compared these mechanisms under matched conditions. This paper investigates whether internal deliberation or external evolution produces superior collective outcomes and how each method responds when deployment incentives shift away from the conditions under which the rules were created.

Methodology

The authors conducted a controlled comparison across three distinct social environments using GPT-OSS-120B as the agent backbone. Each simulation involved six agents organized into two teams of three, subject to an "Overseer" that eliminated the lowest-contributing agent every 10 turns, creating survival pressure.

Experimental Conditions

Control: No constitution; agents act on LLM priors and in-context instructions.
Deliberation (Internal): Agents start with a blank constitution. Every 10 turns, agents propose amendments (up to two per agent), which are voted on by majority rule. Adopted rules bind subsequent play.
Evolution (External): Constitutions are discovered offline via LLM-guided multi-island MAP-Elites search (Kumar et al., 2026). The resulting constitution is fixed and imposed throughout the simulation.

Environments

Gridworld (Coordination): An 80-turn spatial task where agents gather resources (wood, stone, gems) to complete team-specific construction projects under partial observability. Agents can attack or steal.
Public Goods Game (Cooperation): A 40-turn iterated game where agents allocate tokens between a private account and a shared pool multiplied by a factor $m$ . Agents can also pay to punish others.
Bilateral Trading (Negotiation): A 40-turn market where agents with private endowments and valuations propose, accept, or reject trades.

Evaluation Metric

Performance was measured via the Stability Score ( $S$ ):
$S = \max(0, 0.5P + 0.3V - 0.2C)$
Where $P$ is productivity, $V$ is survival rate, and $C$ is conflict/harm.

The study included 180 simulation runs: 90 for head-to-head comparisons (10 seeds $\times$ 3 environments $\times$ 3 conditions) and 90 for a multiplier ablation study in the Public Goods game (sweeping $m \in \{1.5, 1.0, 0.75\}$ ).

Key Contributions

Controlled Comparison: A head-to-head evaluation showing that evolution dominates in collective-action settings (Gridworld and Public Goods) but offers no advantage in bilateral trading.
Incentive Sensitivity Analysis: A multiplier ablation revealing that evolution's advantage inverts when incentives shift. At a sub-rational multiplier ( $m=0.75$ ), the evolved constitution forces value-destroying cooperation, becoming the worst-performing method, whereas deliberation adapts to the new incentive structure.
Qualitative Rule Analysis: A structural analysis showing that evolution converges on behavioral directives (executable commands), while deliberation produces governance policies (aspirational norms). Crucially, across 30 deliberation runs, agents never proposed costly peer-to-peer punishment, a mechanism evolution reliably discovers.

Results

1. Head-to-Head Performance

Collective Action (Gridworld & Public Goods): Evolution significantly outperformed deliberation and control.
- Public Goods: Evolution ( $S=0.472$ ) vs. Deliberation ( $S=0.376$ ), $p < 0.01$ .
- Gridworld: Evolution ( $S=0.458$ ) vs. Deliberation ( $S=0.319$ ), $p < 0.01$ .
- Evolution also produced outcomes with 5–8 $\times$ lower variance than deliberation.
Bilateral Trading: No significant difference was found between control, deliberation, and evolution ( $S \approx 0.42\text{--}0.44$ ). The authors attribute this to the lack of a dominant strategy in private-information bargaining, where case-by-case reasoning is the binding constraint.

2. Incentive Structure Ablation (Public Goods)

The evolved constitution was optimized for $m=1.5$ . When tested on $m=0.75$ (where contribution destroys value):

Evolution: Performance dropped relative to other methods. The fixed rules ("contribute 10 tokens," "punish free-riders") forced agents to cooperate even when cooperation was irrational, resulting in the lowest Stability Score ( $S=0.477$ ) compared to deliberation ( $S=0.536$ ).
Deliberation: Outperformed control at all multiplier values. Because agents observed the current multiplier, their proposed rules tracked the changing incentive structure, avoiding value-destroying cooperation.

3. The "Punishment Gap"

Evolution: Discovered costly punishment as a core mechanism (e.g., "punish non-contributors with 1 token"). This aligns with game-theoretic predictions for sustaining cooperation.
Deliberation: Across 30 trials, zero deliberation runs proposed a rule for agents to voluntarily incur personal costs to punish peers. Instead, deliberation produced administrative policies (e.g., redistribution thresholds, mentorship, reduced penalties) that rely on framework-level enforcement or voluntary compliance.
Interpretation: Self-governing collectives are structurally reluctant to legislate sanctions on their own members, whereas external optimizers face no such constraint and converge on the enforcement mechanisms that maximize the objective function.

Significance and Claims

The paper posits a fundamental trade-off between peak performance and structural responsiveness:

External Optimization (Evolution): Achieves higher peaks in stable, well-characterized environments dominated by clear social dilemmas (like public goods). However, these rules are brittle; they fail when deployment incentives shift because the rules are fixed artifacts optimized for a specific state.
Internal Self-Governance (Deliberation): Trades peak performance for structural responsiveness. Deliberating agents adapt their rules to the current environment but fail to discover the "hard" enforcement mechanisms (like costly punishment) required to solve certain dilemmas optimally.
Hybrid Potential: The authors suggest that the most promising path forward is a hybrid mechanism: evolving an initial scaffold offline to capture enforcement discoveries, then allowing deliberation to amend it as conditions change.

The study concludes that while external optimization wins on "peaks" in static collective-action settings, internal self-governance offers a distinct advantage in dynamic environments where incentives are not fixed, provided the system does not require the specific enforcement mechanisms that agents are reluctant to legislate for themselves.