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

MAKE IMPACT

Jacob Eisenstein , Chirag Nagpal , Alekh Agarwal , Ahmad Beirami , Alex D’Amour , Dj Dvijotham , Adam Fisch , Katherine Heller , Stephen Pfohl , Deepak Ramachandran , Peter J. Shaw , Jonathan Berant · YOU? · · 2023 · Open Access · · DOI: https://doi.org/10.48550/arxiv.2312.09244

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.

Type

preprint

Language

en

Landing Page

http://arxiv.org/abs/2312.09244

PDF

https://arxiv.org/pdf/2312.09244

OA Status

green

Cited By

2

Related Works

10

OpenAlex ID

https://openalex.org/W4389821477

Raw OpenAlex JSON

OpenAlex ID

https://openalex.org/W4389821477

Canonical identifier for this work in OpenAlex

DOI

https://doi.org/10.48550/arxiv.2312.09244

Digital Object Identifier

Title

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

Work title

Type

preprint

OpenAlex work type

Language

en

Primary language

Publication year

2023

Year of publication

Publication date

2023-12-14

Full publication date if available

Authors

Jacob Eisenstein, Chirag Nagpal, Alekh Agarwal, Ahmad Beirami, Alex D’Amour, Dj Dvijotham, Adam Fisch, Katherine Heller, Stephen Pfohl, Deepak Ramachandran, Peter J. Shaw, Jonathan Berant

List of authors in order

Landing page

https://arxiv.org/abs/2312.09244

Publisher landing page

PDF URL

https://arxiv.org/pdf/2312.09244

Direct link to full text PDF

Open access

Yes

Whether a free full text is available

OA status

green

Open access status per OpenAlex

OA URL

https://arxiv.org/pdf/2312.09244

Direct OA link when available

Concepts

Computer science, Inference, Generalization, Artificial intelligence, Reinforcement learning, Herding, Exploit, Machine learning, Mathematics, Forestry, Geography, Computer security, Mathematical analysis

Top concepts (fields/topics) attached by OpenAlex

Cited by

2

Total citation count in OpenAlex

Citations by year (recent)

2025: 1, 2024: 1

Per-year citation counts (last 5 years)

Related works (count)

10

Other works algorithmically related by OpenAlex

Full payload