Reproduce “Disambiguating Moral Circuits and Moral Features in Language Models: Moral content, normative judgment or cue words? Evidence from 14 models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Davide Liga, Liuwen Yu, Matteo Magnini
The authors test whether attention heads identified using moral prompts are specific to moral judgments or contribute to evaluation. They compare moral, other normative, and descriptive phrase pairs, with and without “morally,” across 14 open-weight models. The located heads have similar normalized effects on moral and normative pairs; this supports shared evaluation machinery, not the absence of moral-specific computation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.