Reproduce “Mood Matters: How Syntactic Sensitivity Undermines Safety Alignment” 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
Alina Klerings, Jannik Brinkmann, Heiner Stuckenschmidt, Simone Paolo Ponzetto
The authors test whether safety refusals generalize across syntax by paraphrasing 100 English harmful prompts into seven grammatical forms and evaluating ASR@10 across 16 instruction-tuned models. They use sparse autoencoder feature steering on Gemma-2 2B to test causality and examine post-training data. The worst variant raises ASR@10 by 15%–71%; findings are limited to tested prompts and interventions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.