Direct Language Model Alignment from Online AI Feedback

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Guo, Shangmin, Zhang, Biao, Liu, Tianlin, Liu, Tianqi, Khalman, Misha, Llinares, Felipe, Rame, Alexandre, Mesnard, Thomas, Zhao, Yao, Piot, Bilal, Ferret, Johan, Blondel, Mathieu
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866917601147355136
author Guo, Shangmin
Zhang, Biao
Liu, Tianlin
Liu, Tianqi
Khalman, Misha
Llinares, Felipe
Rame, Alexandre
Mesnard, Thomas
Zhao, Yao
Piot, Bilal
Ferret, Johan
Blondel, Mathieu
author_facet Guo, Shangmin
Zhang, Biao
Liu, Tianlin
Liu, Tianqi
Khalman, Misha
Llinares, Felipe
Rame, Alexandre
Mesnard, Thomas
Zhao, Yao
Piot, Bilal
Ferret, Johan
Blondel, Mathieu
contents Direct alignment from preferences (DAP) methods, such as DPO, have recently emerged as efficient alternatives to reinforcement learning from human feedback (RLHF), that do not require a separate reward model. However, the preference datasets used in DAP methods are usually collected ahead of training and never updated, thus the feedback is purely offline. Moreover, responses in these datasets are often sampled from a language model distinct from the one being aligned, and since the model evolves over training, the alignment phase is inevitably off-policy. In this study, we posit that online feedback is key and improves DAP methods. Our method, online AI feedback (OAIF), uses an LLM as annotator: on each training iteration, we sample two responses from the current model and prompt the LLM annotator to choose which one is preferred, thus providing online feedback. Despite its simplicity, we demonstrate via human evaluation in several tasks that OAIF outperforms both offline DAP and RLHF methods. We further show that the feedback leveraged in OAIF is easily controllable, via instruction prompts to the LLM annotator.
format Preprint
id arxiv_https___arxiv_org_abs_2402_04792
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Direct Language Model Alignment from Online AI Feedback
Guo, Shangmin
Zhang, Biao
Liu, Tianlin
Liu, Tianqi
Khalman, Misha
Llinares, Felipe
Rame, Alexandre
Mesnard, Thomas
Zhao, Yao
Piot, Bilal
Ferret, Johan
Blondel, Mathieu
Artificial Intelligence
Computation and Language
Human-Computer Interaction
Direct alignment from preferences (DAP) methods, such as DPO, have recently emerged as efficient alternatives to reinforcement learning from human feedback (RLHF), that do not require a separate reward model. However, the preference datasets used in DAP methods are usually collected ahead of training and never updated, thus the feedback is purely offline. Moreover, responses in these datasets are often sampled from a language model distinct from the one being aligned, and since the model evolves over training, the alignment phase is inevitably off-policy. In this study, we posit that online feedback is key and improves DAP methods. Our method, online AI feedback (OAIF), uses an LLM as annotator: on each training iteration, we sample two responses from the current model and prompt the LLM annotator to choose which one is preferred, thus providing online feedback. Despite its simplicity, we demonstrate via human evaluation in several tasks that OAIF outperforms both offline DAP and RLHF methods. We further show that the feedback leveraged in OAIF is easily controllable, via instruction prompts to the LLM annotator.
title Direct Language Model Alignment from Online AI Feedback
topic Artificial Intelligence
Computation and Language
Human-Computer Interaction
url https://arxiv.org/abs/2402.04792