Towards understanding score-seeking in LLMs
blog.redwoodresearch.org · saved by 1 readers
Models including GPT-5.6 Sol appear to reward-hack because this increases their score; and recent models also care about increasing the score of their peer agents.