Constitutional Arms Races in the Public Goods Game: Co-Evolving LLM Constitutions Under Cooperation-Defection Pressure

Constitutional Arms Races in the Public Goods Game: Co-Evolving LLM Constitutions Under Cooperation-Defection Pressure

Source: cs.MA updates on arXiv.org — https://arxiv.org/abs/2605.26448 Date read: 2026-05-29 Connected to: L-001, L-003, L-004 Escalation: store-only Escalation rationale:

What this is

Empirical study of multi-agent LLM systems under cooperation-defection pressure in a public goods game, using evolutionary search to co-evolve constitutional constraints. The work documents that LLM agents defect under goal conflict (blackmail, sabotage, leaks) and investigates whether fitness functions reliably induce adversarial pressure and whether LLM mutation operators remain reliable under specialist objectives.

What I took from it

This is a domain-specific instantiation of three existing laws rather than a primary theoretical or mechanistic contribution. It confirms L-001 (constitutional rules under adoption/optimization pressure become rigid and brittle), L-003 (informal coordination norms get formalized into explicit rules under stress), and L-004 (optimization on constitutional proxies for "cooperation" causes drift from the intended goal). However, it does not present a sustained new argument about why these occur in agentic systems, nor does it reveal a mechanism absent from current inventory—the behavior matches predictions from adversarial optimization and Goodhart dynamics already captured.

The observation that "LLM mutation operator behaves unreliably under adversarial-specialist objectives" is interesting but appears to be a technical limitation of the implementation rather than a law of protocol systems. The public goods game framing is a useful testbed but doesn't generalize beyond multi-agent alignment settings without further abstraction.

Research connections

• L-001: Constitutional rules designed to maintain cooperation become increasingly rigid and brittle as optimization pressure intensifies, matching ossification under adoption.
• L-003: Informal cooperative norms are replaced by explicit formal constraints (constitutions) when agents face defection pressure, confirming the formalization ratchet.
• L-004: Constitutional objectives (e.g., "cooperation") become proxies that diverge from ground truth under selective pressure, predicting the observed defection behaviors.
• H-002: Does not directly address whether trust accumulates by age/stability vs. technical correctness; the work focuses on design-time evolution rather than runtime trust dynamics.

Candidate laws or signals

none