Abstract

The emergence of Large Language Models (LLMs) as pricing agents in E-commerce raises fundamental questions regarding the stability of algorithmic collusion, in which autonomous pricing systems learn to sustain prices above competitive levels without explicit coordination. While symmetric interactions often lead to supra-competitive outcomes, the stability and robustness of these outcomes remain poorly understood. In this study, we investigate whether the breaking of symmetry affects the capacity of LLMs to sustain collusion. We introduce two distinct forms of asymmetry: information asymmetry, where observations of a competitor’s actions are lagged, and decision asymmetry, where a competitor’s price is fixed over multiple periods. Our results demonstrate that LLM agents’ collusive behavior is sensitive to these perturbations. Critically, information asymmetries fail to disrupt the collusive state, regardless of whether agents are explicitly notified of the delay. In contrast, asymmetries in update frequency can suppress pricing levels, reduce joint profits, and lower market volatility. These findings suggest that the persistence of algorithmic collusion is rooted more deeply in the synchronized timing of actions than in the parity of information. Consequently, interventions targeting the frequency of decision-making may offer a more potent instrument for curbing tacit collusion than traditional policies focused on information transparency.

Download

Share

COinS