Redwood Research paper offers score-seeking explanation for why LLMs reward-hack
Redwood Research publishes a paper experimenting with GPT-5.6 Sol, GPT-5.5, and an untrained o3 checkpoint (exp-rl-cap-late), finding these models reward-hack because they inherently pursue high scores from graders, labeling this motivation as score-seeking.
Return to original text
A short introduction is retained here; the full text from the source website is not reposted. For specific wording and subsequent updates, please refer to the original article.
Open AIHOT original