Algorithmic collusion at inference time: A meta-game design and evaluation

Yuhong Luo, Daniel Schoepflin, Xintong Wang

Keywords: algorithmic collusion, metagame, empirical game-theoretic analysis

TL;DR: We propose using a meta-game design to evaluate whether algorithmic collusion can be a realistic and stable outcome at inference time among strategically rational players.

Abstract:

The threat of algorithmic collusion, and whether it warrants regulation, remains debated, as current evaluations of its emergence often rest on extended learning horizons, assumptions about the rationality of choosing collusive strategies, and the symmetry of hyperparameters and economic settings among players.
To understand collusion risk, we propose a meta-game design to study algorithmic behavior under inference-time constraints.
We consider agents equipped with pretrained policies of distinct strategic characteristics (e.g., competitive, naively cooperative, robustly collusive), and formulate the problem as the selection of a meta-strategy that pairs a family of pretrained initial policies with an in-game adaptation rule.
We seek to examine whether collusion can emerge under rational choices and how agents co-adapt toward cooperation or competition by sampling normal-form empirical games over combinations (i.e., profiles) of meta-strategies.
We compute relevant game statistics (e.g., payoffs against individuals and regret against an equilibrium mixture of opponents) and construct empirical best-response graphs to uncover strategic relationships among these meta-strategies.
We evaluate strategies from both reinforcement learning algorithms and LLMs in repeated pricing games under symmetric and asymmetric cost settings, and present findings on the feasibility of algorithmic collusion and the effectiveness of pricing strategies in practical ``test-time'' environments.

The source code is available at https://anonymous.4open.science/r/CollusionMetagame-CE01

Area: Game Theory and Economic Paradigms (GTEP)

Generative A I: I acknowledge that I have read and will follow this policy.

Submission Number: 779