Generate multiple chain-of-thought answers, then vote or aggregate—trades cost for reliability on some reasoning tasks.
Self-consistency pairs with chain-of-thought: sample diverse reasoning paths at non-zero temperature, then choose the majority answer. It can reduce some errors but is not a proof: all paths may share a wrong assumption.
Operations
Budget tokens and latency explicitly—this is an n× multiplier on generation.