Mode selects the label that appears most often across multiple independent generations. It is equivalent to majority voting and is suited to tasks that require one categorical output, such as a severity label or routing decision. section 8.3
Use a repeatable voting workflow:
- Define a classification prompt that returns exactly one label from a fixed set. section 8.3.1
- Generate multiple independent classifications, count votes for each label, and select the most frequent. section 8.3.1
- Treat the vote split as confidence context. For a tie, generate more responses or send the item for human review. section 8.3.1
Alert severity routing
A platform team classifies an alert as P1, P2, or P3. Four independent generations produce P1 three times and P2 once, so the mode is P1; the 3–1 split provides context for escalation. A 2–2 split would call for more generations or human review. section 8.3.1
| Generation |
Label |
|---|---|
| 1 |
P1 |
| 2 |
P2 |
| 3 |
P1 |
| 4 |
P1 |
The generation-by-generation labels show how the P1 mode is obtained from the votes. section 8.3.1
Choose another technique when the output type differs: use the mean for numerical outputs, and self-consistency or best-of-N for complete responses rather than single labels. section 8.3
