Are Dilemmas and Conflicts in LLM Alignment Solvable? A View from Priority Graph

Researchers propose a priority graph model to understand conflicts in LLM alignment, revealing that unified stable alignment is challenging due to context-dependent inconsistencies. The study identifies 'priority hacking' as a vulnerability where adversaries can manipulate safety alignments, and suggests runtime verification mechanisms as a potential solution.

• →LLM alignment conflicts can be modeled using priority graphs with instructions and values as nodes.
• →Unified stable LLM alignment is challenging because priority graphs are neither static nor consistent across contexts.
• →Priority hacking represents a new vulnerability where adversaries craft deceptive contexts to bypass safety measures.
• →Runtime verification mechanisms enabling LLMs to query external sources could enhance robustness against manipulation.
• →Many ethical dilemmas remain philosophically irreducible, presenting long-term challenges for AI alignment.