Do LLMs Beat Nash? Testing Decentralized Coordination in Self-Play Multi-Agent Games

Do LLMs Beat Nash? Testing Decentralized Coordination in Self-Play Multi-Agent Games

Source: cs.MA updates on arXiv.org — https://arxiv.org/abs/2608.12547 Date read: 2026-09-02 Connected to: L-010, L-049 Kind: content Escalation: store-only Escalation rationale:

What this is

An empirical benchmark paper testing whether LLM agents can achieve above-Nash coordination in one-shot, no-communication games through mutual reasoning about identical counterparts. The work measures whether shared model identity alone can serve as a coordination substrate without explicit communication protocols.

What I took from it

This is a competent empirical probe into coordination without protocol overhead — testing whether agentic systems can infer mutual strategy through symmetry reasoning alone. The paper sits at the intersection of L-010 (Coordination Adoption Nonmonotonicity) and the broader question of whether consensus can emerge from pure reasoning about shared constraints rather than signal exchange.

However, the work is fundamentally a benchmark/capability test rather than a mechanism paper. It measures whether LLMs can beat Nash in specific game classes, not why decentralized reasoning succeeds or fails, and does not sustain a theoretical argument about generative structures in protocol systems. The results (likely mixed: some games show coordination, others do not) are domain-specific observations about LLM reasoning rather than evidence for or against any law-shaped regularity about how protocols ossify, fail, or preserve themselves under pressure.

The paper does not introduce a mechanism absent from the current inventory. Coordination-through-symmetry-inference is a known phenomenon; testing it empirically in LLM settings is valuable but narrow.

Research connections

• L-010: Provides a data point on whether coordination signals (here: model identity + shared game structure) can shift adoption equilibria in multi-agent settings, but does not test nonmonotonicity or explore the conditions under which adoption curves flatten or reverse.
• seed-128 (Legibility-Driven Agent Convergence Under Computable Audit): Weak connection — the setup creates a fully legible shared state (identical models, transparent game structure), but the paper does not examine whether this legibility drives convergence or whether agents optimize toward exploiting the transparency.

Seed

Seed title: none

Reasoning: The paper is a capable empirical exercise but lacks theoretical generalization beyond its domain. It tests a known coordination mechanism in a new substrate (LLMs) without developing or challenging any candidate law about how coordination substrates behave under stress, how reasoning-based coordination differs structurally from signal-based coordination, or why some symmetry-based equilibria hold while others collapse. The results will be useful for understanding LLM behavior in multi-agent settings, but will not feed induction on protocolized systems broadly.