Reproduce “Language Models Act on Hidden Valence” 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

Cameron Berg, Caspar Kaiser

The authors test whether internal activation steering can affect a language model’s choices between two arbitrary zones. They apply positive or negative valence-related directions while rebuilding the model’s stored history, then switch steering off and hold the visible text fixed. Choice shifts in five of seven models, though two show exceptions; this does not establish subjective experience.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.