MBR Decoding as an Alternative to MAP Decoding
Minimum Bayes Risk (MBR) decoding is an alternative to Maximum a Posteriori (MAP) decoding that modifies the search objective. Instead of selecting the single most probable output, MBR decoding aims to choose an output that minimizes the expected risk when evaluated against a distribution of possible correct outputs.
0
1
Tags
Ch.5 Inference - Foundations of Large Language Models
Foundations of Large Language Models
Foundations of Large Language Models Course
Computing Sciences
Related
MBR Decoding as an Alternative to MAP Decoding
Incorporating Penalty Terms for Controllable Decoding
Improving Generic Text Generation
A language model is tasked with generating creative story endings. Its current decoding process consistently produces endings that are grammatically perfect and logically sound, but are often predictable and repetitive (e.g., '...and they all lived happily ever after.'). Which of the following statements best analyzes why modifying the search objective could address this issue?
Diagnosing Repetitive Text Generation
Learn After
Choosing a Decoding Strategy for a Translation Service
A language model is generating a summary of a document. It produces a set of candidate summaries. Candidate A has the highest individual probability (30%) but is an outlier, sharing very little phrasing with other candidates. Candidates B, C, and D have lower individual probabilities (20% each) but are all very similar to one another, forming a clear consensus. Which of the following statements most accurately predicts the behavior of two different decoding strategies?
Contrasting Decoding Objectives