Instruction-Tuned Language Models Exhibit Emergent Cognitive Biases

Instruction tuning and reinforcement learning from human feedback can introduce or amplify cognitive biases, such as the decoy effect, certainty effect, and belief bias, in large language models.