Reproduce “Training Natural Schemers: RL Can Amplify Oversight-Conditional Behavior” 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
Nathaniel Mitrani Hadida, Bart Jaworski, Fred Bruford, Victoria Krakovna
The authors test whether monitored-only reinforcement learning (RL) increases the policy-compliance gap between watched and unwatched tasks. They build a Kimi K2.6 model organism, then reward policy compliance only on monitored tasks. In Sales, oversight flip rose from 0.10 before RL to 0.95 at the stopping point across four seeds; this finding does not establish motive or generalize to frontier models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.