Layer self-consistency on chain-of-thought for a measured +17.9% on hard reasoning

SurePrompts·medium signal

For non-reasoning models on math and logic tasks, sampling multiple chain-of-thought paths and taking a majority vote (self-consistency) lifts GSM8K accuracy by +17.9% over a single greedy CoT pass. It's the highest-ROI accuracy lever for reasoning-heavy prompts — but skip it entirely for native reasoning models, which already reason internally and gain nothing.