Reproduce “Cultural Value Alignment Via Latent Activation Steering in Large Language 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
Trung Duc Anh Dang, Sarah Masud
The paper tests scenario-based dilemmas and activation steering to reveal and shift preferences in three open models with roughly 3–4 billion parameters. It maps choices from English-language scenarios onto two World Values Survey axes and intervenes during inference without retraining. Scenario probing shifts models toward secular-rational and self-expression values, but steering one axis also shifts the other, with entanglement ratios of 0.7–0.9; real-world behavior is untested.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.