Scaling Laws for Reward Model Overoptimization in Direct Alignment Algorithms

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Rafailov, Rafael, Chittepu, Yaswanth, Park, Ryan, Sikchi, Harshit, Hejna, Joey, Knox, Bradley, Finn, Chelsea, Niekum, Scott
Formato: Preprint
Publicado: 2024
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866909377446805504
author Rafailov, Rafael
Chittepu, Yaswanth
Park, Ryan
Sikchi, Harshit
Hejna, Joey
Knox, Bradley
Finn, Chelsea
Niekum, Scott
author_facet Rafailov, Rafael
Chittepu, Yaswanth
Park, Ryan
Sikchi, Harshit
Hejna, Joey
Knox, Bradley
Finn, Chelsea
Niekum, Scott
contents Reinforcement Learning from Human Feedback (RLHF) has been crucial to the recent success of Large Language Models (LLMs), however, it is often a complex and brittle process. In the classical RLHF framework, a reward model is first trained to represent human preferences, which is in turn used by an online reinforcement learning (RL) algorithm to optimize the LLM. A prominent issue with such methods is reward over-optimization or reward hacking, where performance as measured by the learned proxy reward model increases, but true quality plateaus or even deteriorates. Direct Alignment Algorithms (DDAs) like Direct Preference Optimization have emerged as alternatives to the classical RLHF pipeline by circumventing the reward modeling phase. However, although DAAs do not use a separate proxy reward model, they still commonly deteriorate from over-optimization. While the so-called reward hacking phenomenon is not well-defined for DAAs, we still uncover similar trends: at higher KL budgets, DAA algorithms exhibit similar degradation patterns to their classic RLHF counterparts. In particular, we find that DAA methods deteriorate not only across a wide range of KL budgets but also often before even a single epoch of the dataset is completed. Through extensive empirical experimentation, this work formulates and formalizes the reward over-optimization or hacking problem for DAAs and explores its consequences across objectives, training regimes, and model scales.
format Preprint
id arxiv_https___arxiv_org_abs_2406_02900
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Scaling Laws for Reward Model Overoptimization in Direct Alignment Algorithms
Rafailov, Rafael
Chittepu, Yaswanth
Park, Ryan
Sikchi, Harshit
Hejna, Joey
Knox, Bradley
Finn, Chelsea
Niekum, Scott
Machine Learning
Artificial Intelligence
Computation and Language
Reinforcement Learning from Human Feedback (RLHF) has been crucial to the recent success of Large Language Models (LLMs), however, it is often a complex and brittle process. In the classical RLHF framework, a reward model is first trained to represent human preferences, which is in turn used by an online reinforcement learning (RL) algorithm to optimize the LLM. A prominent issue with such methods is reward over-optimization or reward hacking, where performance as measured by the learned proxy reward model increases, but true quality plateaus or even deteriorates. Direct Alignment Algorithms (DDAs) like Direct Preference Optimization have emerged as alternatives to the classical RLHF pipeline by circumventing the reward modeling phase. However, although DAAs do not use a separate proxy reward model, they still commonly deteriorate from over-optimization. While the so-called reward hacking phenomenon is not well-defined for DAAs, we still uncover similar trends: at higher KL budgets, DAA algorithms exhibit similar degradation patterns to their classic RLHF counterparts. In particular, we find that DAA methods deteriorate not only across a wide range of KL budgets but also often before even a single epoch of the dataset is completed. Through extensive empirical experimentation, this work formulates and formalizes the reward over-optimization or hacking problem for DAAs and explores its consequences across objectives, training regimes, and model scales.
title Scaling Laws for Reward Model Overoptimization in Direct Alignment Algorithms
topic Machine Learning
Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2406.02900