Self-Consistency
Sample several reasoning paths and let them vote — buying accuracy with compute on tasks that have a checkable final answer.
Self-consistency upgrades chain-of-thought with a simple observation: there are many ways to reason wrong but few ways to reason right. Run the same CoT prompt several times at a temperature high enough to vary the reasoning path, collect only the final answers, and take the majority. Erroneous paths tend to scatter — a dropped negative here, a misread constraint there — landing on different wrong answers. Correct paths converge on the same one. The vote filters the semi-random errors out. No new prompt engineering, no fine-tuning: you're spending inference-time compute to buy reliability.
The technique has one hard requirement: a voteable answer. Numbers, labels, multiple-choice letters, yes/no — anything short and comparable. Ten essays can't vote; ten values of '7' can. That's why self-consistency pairs naturally with the answer-extraction contracts from the last lesson — the FINAL: line is exactly what you tally. For open-ended output you need different aggregation: an LLM judge picking the best of N, or clustering similar responses. That's really ensembling, which Module 3 covers. Know which regime your task is in before reaching for votes.
The economics: N samples cost N times the tokens and, run in parallel, one call's latency. Five to ten samples capture most of the gain; returns diminish steeply after that. Self-consistency makes sense for high-stakes answers where being wrong costs more than ten LLM calls — a medical-coding label, a financial calculation, grading that gates something real — and for offline batch work where latency is irrelevant. It's the wrong tool for chat-speed interactions and for easy tasks the model already gets right 98% of the time. And a tie is information: a 4-3-3 split is the model telling you the question is genuinely hard — escalate it.
Implementation notes: temperature matters — at 0 every path is identical and the vote is theater; around 0.7 to 1.0 the paths genuinely diverge. Extract answers programmatically, never by eyeballing. Log the vote distribution, not just the winner: 9-1 and 5-4-1 are very different confidence signals hiding behind identical winners. And note that 2026 reasoning models internalize a cousin of this trick — some spend their thinking budget exploring alternative paths — but cross-sampling and voting remains yours to apply on any model, including them.