A vote fixes scattered errors, and locks in shared ones.
Self-consistency samples N reasoning paths and keeps the most common final answer. This lab computes the exact chance that the vote is right, for any N, any accuracy of one path, and any number of distinct wrong answers.
- One path
- Vote of N paths
- Change, points
- Calls, LLM aggregator
Vote accuracy by number of paths
Exact, ties split at random
What each extra path buys
| N | Vote accuracy | Change from one path | Calls, code vote | Calls, LLM aggregator |
|---|
Each path is an independent draw here. Paths sampled from one model share its blind spots, so a real vote gains less than this curve. Notebook 02a samples three paths and asks a fourth call to aggregate them.