Frame-Conditioned Moral Computation in LLaMA 3.1-8B-Instruct: A Mechanistic Interpretability Audit of Ethical Reasoning

The paper presents an analysis of the LLaMA 3.1-8B-Instruct model's ethical reasoning capabilities through a mechanistic interpretability framework using the Transluce platform. It investigates the model's responses to 54 moral prompts across various scenarios, revealing that the model's ethical reasoning is influenced by the framing of the prompts rather than its ethical capacity itself. This study introduces the concept of Frame-Conditioned Moral Computation, emphasizing the need for Mechanistic Alignment in AI ethics, where the interpretive frame significantly impacts the model's moral conclusions, highlighting the importance of understanding internal computations for practitioners developing ethically aligned AI systems.