Helping or Herding? Reward Model Ensembles Mitigate but do not Eliminate Reward Hacking

• URL: http://arxiv.org/abs/2312.09244v3
• Date: Fri, 16 Aug 2024 23:59:29 GMT
• Title: Helping or Herding? Reward Model Ensembles Mitigate but do not Eliminate Reward Hacking
• Authors: Jacob Eisenstein, Chirag Nagpal, Alekh Agarwal, Ahmad Beirami, Alex D'Amour, DJ Dvijotham, Adam Fisch, Katherine Heller, Stephen Pfohl, Deepak Ramachandran, Peter Shaw, Jonathan Berant,
• Abstract summary: Reward models play a key role in aligning language model applications towards human preferences. A natural mitigation is to train an ensemble of reward models, aggregating over model outputs to obtain a more robust reward estimate. We show that reward ensembles do not eliminate reward hacking because all reward models in the ensemble exhibit similar error patterns.
• Score: 62.146953368613815
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Reward models play a key role in aligning language model applications towards human preferences. However, this setup creates an incentive for the language model to exploit errors in the reward model to achieve high estimated reward, a phenomenon often termed \emph{reward hacking}. A natural mitigation is to train an ensemble of reward models, aggregating over model outputs to obtain a more robust reward estimate. We explore the application of reward ensembles to alignment at both training time (through reinforcement learning) and inference time (through reranking). First, we show that reward models are \emph{underspecified}: reward models that perform similarly in-distribution can yield very different rewards when used in alignment, due to distribution shift. Second, underspecification results in overoptimization, where alignment to one reward model does not improve reward as measured by another reward model trained on the same data. Third, overoptimization is mitigated by the use of reward ensembles, and ensembles that vary by their \emph{pretraining} seeds lead to better generalization than ensembles that differ only by their \emph{fine-tuning} seeds, with both outperforming individual reward models. However, even pretrain reward ensembles do not eliminate reward hacking: we show several qualitative reward hacking phenomena that are not mitigated by ensembling because all reward models in the ensemble exhibit similar error patterns.

Related papers

• R3HF: Reward Redistribution for Enhancing Reinforcement Learning from Human Feedback [25.27230140274847]
  Reinforcement learning from human feedback (RLHF) provides a paradigm for aligning large language models (LLMs) with human preferences. This paper proposes a novel reward redistribution method called R3HF, which facilitates a more fine-grained, token-level reward allocation.
  arXiv  Detail & Related papers  (2024-11-13T02:45:21Z)
• Towards Reliable Alignment: Uncertainty-aware RLHF [14.20181662644689]
  We show that the fluctuation of reward models can be detrimental to the alignment problem. We show that such policies are more risk-averse in the sense that they are more cautious of uncertain rewards. We use this ensemble of reward models to align language model using our methodology and observe that our empirical findings match our theoretical predictions.
  arXiv  Detail & Related papers  (2024-10-31T08:26:51Z)
• Evaluating Robustness of Reward Models for Mathematical Reasoning [14.97819343313859]
  We introduce a new design for reliable evaluation of reward models, and to validate this, we construct RewardMATH. We demonstrate that the scores on RewardMATH strongly correlate with the results of optimized policy and effectively estimate reward overoptimization.
  arXiv  Detail & Related papers  (2024-10-02T16:39:58Z)
• Elephant in the Room: Unveiling the Impact of Reward Model Quality in Alignment [50.21842377409232]
  Despite vital role reward models play in alignment, previous works have consistently overlooked their performance. This work first investigates the quality of the widely-used preference dataset, HH-RLHF, and curates a clean version, CHH-RLHF. Based on CHH-RLHF, we benchmark the accuracy of a broad range of reward models used in previous alignment works, unveiling the unreliability of using them both for optimization and evaluation.
  arXiv  Detail & Related papers  (2024-09-26T04:28:35Z)
• HAF-RM: A Hybrid Alignment Framework for Reward Model Training [51.59246299566669]
  We propose a hybrid alignment framework HaF-RM for reward model training. It offers a principled and effective approach to enhancing the performance and alignment of reward models.
  arXiv  Detail & Related papers  (2024-07-04T23:26:56Z)
• RewardBench: Evaluating Reward Models for Language Modeling [100.28366840977966]
  We present RewardBench, a benchmark dataset and code-base for evaluation of reward models. The dataset is a collection of prompt-chosen-rejected trios spanning chat, reasoning, and safety. On the RewardBench leaderboard, we evaluate reward models trained with a variety of methods.
  arXiv  Detail & Related papers  (2024-03-20T17:49:54Z)
• Transforming and Combining Rewards for Aligning Large Language Models [69.44634017612798]
  A common approach for aligning language models to human preferences is to first learn a reward model from preference data, and then use this reward model to update the language model. We use a log-sigmoid function to transform rewards learned from Bradley-Terry preference models. Experiments aligning language models to be both helpful and harmless using RLHF show substantial improvements over the baseline (non-transformed) approach.
  arXiv  Detail & Related papers  (2024-02-01T16:39:28Z)
• Reward Collapse in Aligning Large Language Models [64.98482888193267]
  We study the phenomenon of textitreward collapse', an empirical observation where the prevailing ranking-based approach results in an textitidentical reward distribution. Our experimental results suggest that our proposed prompt-aware utility functions significantly alleviate reward collapse during the training of reward models.
  arXiv  Detail & Related papers  (2023-05-28T02:12:00Z)
• Scaling Laws for Reward Model Overoptimization [19.93331579503503]
  We study how the gold reward model score changes as we optimize against the proxy reward model using either reinforcement learning or best-of-$n$ sampling. We also study the effect on this relationship of the size of the reward model dataset, the number of reward model and policy parameters, and the coefficient of the KL penalty added to the reward in the reinforcement learning setup.
  arXiv  Detail & Related papers  (2022-10-19T17:56:10Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.