Reproduce “Emotion Concepts Across Model Scale” 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
Sanyam Jain
The authors measure emotion-related directions in activations across 19 open base models from three families. They derive 171 emotion vectors and compare geometry across sizes. Synonym–antonym separation rises from 0.250 at 111M parameters to 0.637 at 1.3B, then is 0.635 at 13B. Geometry largely plateaus around 1B, but not every measure does; these findings concern activations, not subjective feelings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.