Reproduce “No Step Looked Wrong: FadeBench, a Benchmark for Incremental Moral Drift in Small 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
Becky R. Northaven
FadeBench tests whether models choose harmful options more often after a sequence of small decisions than when the final decision is asked alone. It uses four-step ladders about dependents and paired cold controls. For bare gemma-4-E4B with reasoning on, terminal drift was 15% versus under 2% cold, a 13.2-point gap across 17 scenario versions; this result does not establish a general effect across models or causes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.