CoffeeBench Evaluates LLM Agent Capability in Industrial Fault Diagnosis

CoffeeBench introduces a simulated appliance task to test LLM agents on real-world industrial asset management. Findings show that models struggle with raw sensor data alone but succeed when iteratively refining fault-option spaces through diagnostic reasoning.