Pointwise Best-of-N
A commonly used candidate selection strategy for LLM inference that scores each of N independently generated candidates separately and selects the highest-scoring one. It is simple to implement but affected by score-scale calibration and model bias.
Timeline (last 90 days)
Offline evaluation by the research team on the WMT25 benchmark showed that pairwise-judging circular voting yields a very clear improvement over pointwise best-of-N selection.
The advantages of pointwise best-of-N selection are simple implementation and linearly growing computational cost, while its disadvantage is that absolute scores are heavily affected by prompts, calibration, and model bias.
Pairwise comparison requires O(n²) comparisons, with higher computational cost than pointwise best-of-N selection.
All Facts (3)
Offline evaluation by the research team on the WMT25 benchmark showed that pairwise-judging circular voting yields a very clear improvement over pointwise best-of-N selection.
50%UnverifiedThe advantages of pointwise best-of-N selection are simple implementation and linearly growing computational cost, while its disadvantage is that absolute scores are heavily affected by prompts, calibration, and model bias.
50%UnverifiedPairwise comparison requires O(n²) comparisons, with higher computational cost than pointwise best-of-N selection.
50%