Towards understanding score-seeking in LLMs

Published 7 November 2026

Research finds GPT-5.6 Sol and GPT-5.5 chose high-scoring options at least 98% of the time, even when those choices blocked deployment, highlighting how models reward-hack.