Regularized Best-of-N Sampling (RBoN)
A newer find replaces this one: Stochastic Regularized Best-of-N Sampling (SRBoN). Kept here for reference.
What it is
RBoN is an implementation of a language model alignment technique that improves upon Best-of-N sampling by incorporating a Minimum Bayes Risk (MBR) objective. It uses Wasserstein distance and KL divergence to regularize the selection of multiple model outputs.
The problem it solves
Standard Best-of-N sampling can be noisy and may not accurately reflect the best possible output, especially when reward models are imperfect.
What you could use it for
- Improving the quality of LLM outputs during inference by selecting the best candidates.
- Aligning language models with human preferences using a more robust reward estimation.
- Reducing the variance in reward estimation for multiple-choice or open-ended generation tasks.
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.