more thinking, more bias: length-driven position bias in reasoning models
a study finds that longer chain-of-thought reasoning in ai models correlates with increased position bias in multiple-choice questions, challenging the assumption that more thinking reduces shallow biases.