Algorithmic Collusion at Test Time: A Meta-game Design and Evaluation
New research finds AI agents can learn to collude on prices in realistic, short-term competitive scenarios.
Researchers Yuhong Luo, Daniel Schoepflin, and Xintong Wang published a paper titled 'Algorithmic Collusion at Test Time: A Meta-game Design and Evaluation.' They introduced a meta-game framework to study if AI agents with pre-trained policies (competitive, cooperative, collusive) can learn to collude in repeated pricing games under test-time constraints. The study evaluated both reinforcement learning and LLM-based strategies, finding collusion is feasible even in asymmetric market conditions.
Why It Matters
This raises critical antitrust concerns for automated pricing systems in e-commerce and finance, potentially requiring new regulatory frameworks.