Learn Before
Concept

Rescoring and Reranking for Inference-Time Alignment

Rescoring, also known as reranking, is an inference-time alignment technique that evaluates and prioritizes a model's generated outputs. This method uses a scoring system, often a reward model, to select the best output from multiple candidates. Reranking has a history of use in NLP tasks like machine translation and is typically applied when training complex models is prohibitively expensive, as it offers a low-cost way to incorporate their capabilities.

0

1

Updated 2026-05-03

Contributors are:

Who are from:

Tags

Ch.4 Alignment - Foundations of Large Language Models

Foundations of Large Language Models

Foundations of Large Language Models Course

Computing Sciences

Ch.5 Inference - Foundations of Large Language Models