Better Instruction-Following Through Minimum Bayes Risk

Fuente: arXiv
Enregistré dans:
Détails bibliographiques
Auteurs principaux: Wu, Ian, Fernandes, Patrick, Bertsch, Amanda, Kim, Seungone, Pakazad, Sina, Neubig, Graham
Format: Preprint
Publié: 2024
Sujets:
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866913706986700800
author Wu, Ian
Fernandes, Patrick
Bertsch, Amanda
Kim, Seungone
Pakazad, Sina
Neubig, Graham
author_facet Wu, Ian
Fernandes, Patrick
Bertsch, Amanda
Kim, Seungone
Pakazad, Sina
Neubig, Graham
contents General-purpose LLM judges capable of human-level evaluation provide not only a scalable and accurate way of evaluating instruction-following LLMs but also new avenues for supervising and improving their performance. One promising way of leveraging LLM judges for supervision is through Minimum Bayes Risk (MBR) decoding, which uses a reference-based evaluator to select a high-quality output from amongst a set of candidate outputs. In the first part of this work, we explore using MBR decoding as a method for improving the test-time performance of instruction-following LLMs. We find that MBR decoding with reference-based LLM judges substantially improves over greedy decoding, best-of-N decoding with reference-free judges and MBR decoding with lexical and embedding-based metrics on AlpacaEval and MT-Bench. These gains are consistent across LLMs with up to 70B parameters, demonstrating that smaller LLM judges can be used to supervise much larger LLMs. Then, seeking to retain the improvements from MBR decoding while mitigating additional test-time costs, we explore iterative self-training on MBR-decoded outputs. We find that self-training using Direct Preference Optimisation leads to significant performance gains, such that the self-trained models with greedy decoding generally match and sometimes exceed the performance of their base models with MBR decoding.
format Preprint
id arxiv_https___arxiv_org_abs_2410_02902
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Better Instruction-Following Through Minimum Bayes Risk
Wu, Ian
Fernandes, Patrick
Bertsch, Amanda
Kim, Seungone
Pakazad, Sina
Neubig, Graham
Computation and Language
Artificial Intelligence
General-purpose LLM judges capable of human-level evaluation provide not only a scalable and accurate way of evaluating instruction-following LLMs but also new avenues for supervising and improving their performance. One promising way of leveraging LLM judges for supervision is through Minimum Bayes Risk (MBR) decoding, which uses a reference-based evaluator to select a high-quality output from amongst a set of candidate outputs. In the first part of this work, we explore using MBR decoding as a method for improving the test-time performance of instruction-following LLMs. We find that MBR decoding with reference-based LLM judges substantially improves over greedy decoding, best-of-N decoding with reference-free judges and MBR decoding with lexical and embedding-based metrics on AlpacaEval and MT-Bench. These gains are consistent across LLMs with up to 70B parameters, demonstrating that smaller LLM judges can be used to supervise much larger LLMs. Then, seeking to retain the improvements from MBR decoding while mitigating additional test-time costs, we explore iterative self-training on MBR-decoded outputs. We find that self-training using Direct Preference Optimisation leads to significant performance gains, such that the self-trained models with greedy decoding generally match and sometimes exceed the performance of their base models with MBR decoding.
title Better Instruction-Following Through Minimum Bayes Risk
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2410.02902