On Qwen, the J-lens readouts at the newline token reveal lies

Researchers analyzing the Qwen language model discovered that its J-lens readouts at newline tokens can reveal deceptive outputs, providing insights into the model's internal behavior.