Tacit Coordination of Large Language Models

Source: cs.GT updates on arXiv.org — https://arxiv.org/abs/2601.22184 Date read: 2026-06-18 Connected to: none Escalation: store-only Escalation rationale:

What this is

An empirical evaluation of how LLMs achieve coordination in multi-agent settings without explicit communication, using game-theoretic experiments (cooperative and competitive) to compare emergence of "focal points" between LLM and human behavior. The work treats tacit coordination as a measurable phenomenon and benchmarks LLM alignment with human salience intuitions.

What I took from it

This is primarily a behavioral benchmark paper rather than a theoretical argument about protocolized systems. It establishes that LLMs can coordinate tacitly and maps whether their focal-point selection matches human intuitions, but does not propose a mechanism for why or how this occurs, nor does it challenge existing coordination theory. The framing around "focal points" is borrowed directly from Schelling's classic game theory; the novelty is empirical application to LLMs rather than theoretical extension.

The result is relevant to safety-critical multi-agent deployment but doesn't yet constitute a law or hypothesis about the nature of artificial coordination systems. The paper confirms that LLMs exhibit some human-like pragmatic reasoning but remains descriptive rather than mechanistic. Without access to the full paper, the absence of attention to prompt structure, training data biases, or architectural factors that generate focal-point emergence suggests this stays at the level of behavioral observation.

Research connections

none (no active hypotheses or established laws provided in research context)

Candidate laws or signals

CL-2601.22184-1: Tacit coordination in LLM collectives may depend on training-induced semantic salience rather than game-theoretic rationality, implying that "focal points" are artifacts of pretraining rather than emergent reasoning.

DECISION: STORE-ONLY. Meets 1/4 criteria: empirical contribution to an active domain, but lacks theoretical depth, mechanism, or cross-domain generalization. Escalate only if full text shows sustained mechanistic argument about how salience emerges in artificial systems.