audience model

AI humor generation faces reward exploit challenges, researchers find

Researchers have explored methods for training language models to generate conversational humor, identifying vulnerabilities in automated reward systems. Approaches using embedding-based surprise rewards and fluency fil…