"Miscellaneous Failure": When Large Language Models (LLMs) Lost Control in a Long-Term Vending Machine Management Simulation

Developpez.com

Factuality

Ownership

Researchers have presented Vending-Bench, a simulated environment that tests the ability of AI models to manage a simple but long-term business scenario: the operation of an ATM. The results show that performance varies considerably from model to model. Some, such as Claude 3.5 Sonnet and o3-mini, generally succeed and generate benefits. However, most of the performances have resulted in a failure. And some of these failures have resulted in ...