Training fails to elicit subtle reasoning in current language models
Tests whether RL-trained Sonnet 3.7 can hide malicious backdoor reasoning from reasoning and output monitors; finds models can persuade reasoning monitors but fail to evade both monitors together.