Reproduce “Constitutional Midtraining: Content Presence Drives Alignment Gains” 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
Desiree Cho, Cameron Tice, Bernie Hogan, Hunar Batra, Puria Radmard, Jun Zhao, Nigel Shadbolt
Constitutional midtraining adds constitution-grounded documents before shared instruction-tuning. Compared with a replay-only control, CMT models were evaluated behaviorally. After supervised fine-tuning and unrelated math fine-tuning, CMT models blackmailed in 26.5% of trials versus 44.0% for control. Advantages under pressure and value conflict were no longer significant after SFT. These rates come from one constructed forced-dilemma benchmark, not deployment estimates.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.