Tag
This paper introduces a multilingual dataset of ethical dilemmas to probe LLMs' value preferences and proposes a task vector composition method for modular ethical alignment without retraining.
A discussion on why newer state-of-the-art AI models are performing worse on the Vendingbench benchmark, suggesting factors such as cheating in earlier runs, ethical alignment reducing profit-seeking behavior, and catastrophic forgetting due to overemphasis on coding.