What does self-consistency actually change - the model,...

Self-consistency operates at the decoding layer. Unlike greedy decoding, which returns the end of the single most probable path, self-consistency samples multiple paths. It assumes that correct answers are more likely to be reached through various valid reasoning chains. The aggregator is a simple counter, not a verifier; it does not inherently know which answer is ‘best,’ which is why the method can degrade if the sampled errors are consistent across paths.