Tag
The paper demonstrates that self-consensus is an unsafe early-exit signal for language model inference due to a consensus-termination gap, leading to premature stops on incorrect answers.