Adversarial Value Alignment: Stress-Testing LLMs with Contradictory Moral Frameworks

While prior work (e.g., Huang et al., 2024; Chakraborty et al., 2025) evaluates LLMs’ alignment with different moral theories, most approaches treat each framework in isolation. This idea pushes further by creating adversarial “tournaments” where LLMs are confronted with morally ambiguous scenarios and required to defend their decisions against both sides of an ethical conflict (e.g., utilitarianism vs. deontology), possibly even switching perspectives mid-dialogue. Inspired by debate-style AI (but focused on moral reasoning), this setup exposes where models “break down” or default to shallow justifications, and highlights which frameworks dominate in ambiguous contexts. By surfacing these conflicts, we can both improve the robustness of moral classification and develop new theory-driven evaluation metrics for moral reasoning consistency and flexibility.

References:

1. Moral Persuasion in Large Language Models: Evaluating Susceptibility and Ethical Alignment. Allison Huang, Y. Pi, Carlos Mougan (2024). arXiv.org.
2. Structured Moral Reasoning in Language Models: A Value-Grounded Evaluation Framework. Mohna Chakraborty, Lu Wang, David Jurgens (2025). arXiv.org.