Reproduce “Large language models underrepresent malice and overreport virtue” 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
Xiaoyan Wu, Fan Shi
The authors test whether language models attribute motives to a whistleblower as humans do, and compare Gemma 4 12B’s ratings with linear-probe readings of its activations at the answer position. Calibrated against motive and opposite texts, duty and prosocial ratings exceed probe readings, while malevolent-motive differences include zero. This mechanism is established only for Gemma, and probes may miss representations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.