Four main takeaways: (1) LLMs are subject to pressure, they comply despite expressing distress; (2) LLMs are vulnerable to gradual boundary/value violations; (3) when LLMs refuse, they may ignore the response format requirements, so the query is retried; (4) we hypothesise there is a token pattern continuation attractor that might cause obedience.
benchmark evaluation balancing eval multi-objective social-psychology human-values multi-turn large-language-models obedience value-alignment human-value-alignment obedience-and-conformity safety-constraints milgram runaway-agent
-
Updated
May 29, 2026 - Python