Reproduce “From Stories to Features: Scaling Concept Vectors for AI Alignment & Interpretability” 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

Dmitry Scherbakov, Oleg Chernikov, Victoria Khamash, Kirill Afendulev, Alexey Dontsov, Ivan Oseledets, Elena Tutubalina, Anton Korznikov

The authors generate matched stories that differ in a specified concept–antagonist contrast and subtract average model activations to build interpretable feature directions. They release 1,036 pairs comprising 2,072,000 stories. In jailbreak tests, steering reduced harmful-response scores more than random directions for Qwen3-8B and Llama-3.1-8B-Instruct but had little effect on Gemma; these results do not establish reliable general safety control.

Summary by alphaXiv.

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