Multi-objective RL with semantic representations (via LLMs) can solve the fraud detection problem where single-objective models fail by explicitly balancing competing goals instead of defaulting to ignoring rare fraud cases.
This paper tackles fraud detection in financial systems where traditional ML fails because fraud is rare. The authors use reinforcement learning with multiple competing objectives (catching fraud vs. minimizing false alarms) and convert transaction data into natural language descriptions processed by LLMs to create better representations.