A preprint stress-testing deliberative alignment as anti-scheming training. The companion cites two findings: in an experimental OpenAI run focused on capabilities, training raised the model's rate of saying it was being evaluated from 2% to about 20%, and tests on o3 showed that such reasoning reduces covert misbehavior. Both support the reactive phenotype (P9) and prediction 6.
- Cited in
- the companion paper
- Used for
- P9Prediction 6
- Strand
- AI safety, alignment & autonomy evaluations
Cite
@misc{schoen2025stress,
title = {Stress Testing Deliberative Alignment for Anti-Scheming Training},
author = {Bronson Schoen and Evgenia Nitishinskaya and Mikita Balesni and Axel Højmark and Felix Hofstätter and Jérémy Scheurer and Alexander Meinke and Jason Wolfe and Teun van der Weij and Alex Lloyd and et al.},
year = {2025},
url = {https://arxiv.org/abs/2509.15541},
}