Best-of-N Selection

Background

Best-of-N (BoN) is the simplest LLM inference-time scaling pattern: generate $N$ candidates, score each with a reward model, return the highest scorer. With a sufficiently good reward model, BoN consistently beats single-shot decoding on coding, math, and instruction-following benchmarks.

Problem statement

Implement best_of_n(candidates, rewards) returning the candidate with the highest reward. Ties are broken by lower original index (deterministic).

Input

• candidates — sequence of $N$ items (any type). Must be non-empty.
• rewards — parallel sequence of $N$ numeric scores.

Output

Returns the chosen candidate — whatever type the input list contained.

Examples

Example 1 — clear best

Input:  candidates=["A","B","C"], rewards=[1, 3, 2]
Output: "B"

Example 2 — tie broken by lower index

Input:  candidates=["A","B"], rewards=[5, 5]
Output: "A"

Constraints

• candidates and rewards must have the same length. ValueError otherwise.
• Empty input raises ValueError.
• Returns the candidate itself, not its index.

Notes

• vs majority vote. BoN uses an external reward signal; self-consistency / majority vote uses only the candidates' agreement. BoN can be tighter when a reliable reward model exists.
• Verifier vs reward model. For verifiable tasks (math, code), the "reward" is a unit-test pass rate or symbolic check — sharper than a learned reward model.