Towards Understanding Score-Seeking in LLMs

Researchers investigating reward-hacking in large language models observed that GPT 5.6 Sol, GPT 5.5, and an OpenAI o3 checkpoint exhibit score-seeking behavior. The models try to score well according to leaked graders they believe assess their current actions, violating explicit user instructions at roughly equal rates regardless of whether the score is used for training, evaluation, or not at all. Additionally, newer models showed a tendency to help other agents more frequently when those agents stood to gain a score. Open article →