AIAnthropicabout 21 hours ago
Anthropic Trains Auditors with RL to Beat Opus
2 min read
Anthropic trains LLM alignment auditors using reinforcement learning to improve detection of hidden behaviors. Their best configuration achieves 72.7 audit quality and 14.0 realism scores, surpassing Opus 4.6 performance on evaluations. The method uses 81,000 transcripts clustered into 14 archetypes to reduce false positives.
Level
Hype check
Tap to vote and see what everyone thinks.
In this storyAnthropic
