Why does self-consistency degrade gracefully when a tre...

The difference lies in the aggregation mechanism. Self-consistency is additive: if one of five samples is wrong, the majority vote is still likely to be correct. Tree search is multiplicative: if a scorer incorrectly prunes a branch, that error propagates through the remaining search depth. Because tree pruning is irreversible, the probability of keeping the correct path alive decreases exponentially with each step, making it much more sensitive to scoring errors than a simple voting mechanism.