Google

The set of $$N$$ output sequences, $$\{\hat{\mathbf{y}}_1, ..., \hat{\mathbf{y}}_N\}$$, in Best-of-$$N$$ (BoN) sampling is formally defined as the top $$N$$ sequences that maximize the conditional probability $$\Pr(\mathbf{y}|\mathbf{x})$$. This selection is expressed using the $$\mathrm{argTopN}$$ operation, which returns the $$N$$ outputs that yield the highest values for the function:

$$\{\hat{\mathbf{y}}_1, ..., \hat{\mathbf{y}}_N\} = \mathop{\mathrm{argTopN}}_{\mathbf{y}} \left[ \Pr(\mathbf{y}|\mathbf{x}) \right]$$

Mathematical Formulation of BoN Sampling

Consider the formal definition for selecting a set of N output sequences based on an input sequence, x:

$$\left\{\hat{\mathbf{y}}_1, ..., \hat{\mathbf{y}}_N\right\} = \underset{\mathbf{y}}{\text{argTopN}} [\text{Pr}(\mathbf{y}|\mathbf{x})]$$

In your own words, explain what this mathematical expression represents. Specifically, describe the function of the `Pr(y|x)` component and the `argTopN` operation in this context.

Interpreting the BoN Sampling Formula

In the mathematical formulation of a sampling method where the top N outputs are selected based on their likelihood given an input `x`, the expression `argTopN_y [Pr(y|x)]` returns the N highest conditional probability values.

A language model processes an input prompt `x` and considers four potential output sequences: `y1`, `y2`, `y3`, and `y4`. The model calculates the conditional probability `Pr(y|x)` for each sequence as follows:

*   `Pr(y1|x) = 0.45`
*   `Pr(y2|x) = 0.15`
*   `Pr(y3|x) = 0.85`
*   `Pr(y4|x) = 0.60`

If the model employs a selection method that chooses the top 2 sequences that maximize this probability, which set of sequences will be the final output?

Learn Before

Related