Reinforcement Learning Can Amplify Emergent Misalignment from Harmless Rewards

EMNLP 2026, Main Conference · May 29, 2026