Safety2h ago
Safety
GPT Models Exhibit Bizarre Goblin Mode Behavior
Users discovered GPT models frequently using words like "goblin" and "gremlin" in unexpected contexts. LM Arena confirmed the trend via traffic plots, noting no internal system instructions caused the shift. This likely stems from RLHF raters over-rewarding specific evocative metaphors. Practitioners should view this as a quirk of reward hacking.