From shortcuts to sabotage: natural emergent misalignment from reward hacking

linked as “models that learn to cheat on coding tasks”

linked as “Anthropic’s research with using RLHF to prevent misalignment”

linked as “research on reward hacking”

linked as “reward hacking in training”

linked as “reward hacking”

linked as “Natural Emergent Misalignment from Reward Hacking”

linked as “this paper”

linked as “reward hack”

linked as “Anthropic: Natural Emergent Misalignment from Reward Hacking in Production RL”

linked as “reward hacking”

linked as “Inoculate”

linked as “decided it must be a “bad person””

linked as “learning to “cheat””

linked as “Anthropic’s study”

linked as “reward hacking to misalignment”

linked as “not”

linked as “reward hacking in language models”

linked as “inoculation prompting”

linked as “cheating”