Large language models are increasingly deployed in settings where moral and legal judgments matter, yet how they represent different styles of moral reasoning remains largely unexplored. A new preprint on arXiv investigates whether a model's internal activations carry separable signatures for rule-based versus consequence-based reasoning, moving beyond surface-level output analysis.
The authors ask whether a model's differential processing of these reasoning styles is reflected in its internal representations. By probing the model's hidden states, the work aims to determine if moral reasoning styles are dissociable at the representational level, which could have implications for interpretability and alignment.
Because the abstract is truncated, the specific methods and findings are not fully available in the source. The preprint's significance lies in its framing: understanding moral reasoning in LLMs requires looking inside the model, not just at its answers. No other sources were provided, so this summary reflects only the single preprint's stated aims.