IDEAS home Printed from https://ideas.repec.org/p/arx/papers/2501.06248.html
   My bibliography  Save this paper

Utility-inspired Reward Transformations Improve Reinforcement Learning Training of Language Models

Author

Listed:
  • Roberto-Rafael Maura-Rivero
  • Chirag Nagpal
  • Roma Patel
  • Francesco Visin

Abstract

Current methods that train large language models (LLMs) with reinforcement learning feedback, often resort to averaging outputs of multiple rewards functions during training. This overlooks crucial aspects of individual reward dimensions and inter-reward dependencies that can lead to sub-optimal outcomes in generations. In this work, we show how linear aggregation of rewards exhibits some vulnerabilities that can lead to undesired properties of generated text. We then propose a transformation of reward functions inspired by economic theory of utility functions (specifically Inada conditions), that enhances sensitivity to low reward values while diminishing sensitivity to already high values. We compare our approach to the existing baseline methods that linearly aggregate rewards and show how the Inada-inspired reward feedback is superior to traditional weighted averaging. We quantitatively and qualitatively analyse the difference in the methods, and see that models trained with Inada-transformations score as more helpful while being less harmful.

Suggested Citation

  • Roberto-Rafael Maura-Rivero & Chirag Nagpal & Roma Patel & Francesco Visin, 2025. "Utility-inspired Reward Transformations Improve Reinforcement Learning Training of Language Models," Papers 2501.06248, arXiv.org, revised Feb 2025.
  • Handle: RePEc:arx:papers:2501.06248
    as

    Download full text from publisher

    File URL: http://arxiv.org/pdf/2501.06248
    File Function: Latest version
    Download Restriction: no
    ---><---

    More about this item

    NEP fields

    This paper has been announced in the following NEP Reports:

    Statistics

    Access and download statistics

    Corrections

    All material on this site has been provided by the respective publishers and authors. You can help correct errors and omissions. When requesting a correction, please mention this item's handle: RePEc:arx:papers:2501.06248. See general information about how to correct material in RePEc.

    If you have authored this item and are not yet registered with RePEc, we encourage you to do it here. This allows to link your profile to this item. It also allows you to accept potential citations to this item that we are uncertain about.

    We have no bibliographic references for this item. You can help adding them by using this form .

    If you know of missing items citing this one, you can help us creating those links by adding the relevant references in the same way as above, for each refering item. If you are a registered author of this item, you may also want to check the "citations" tab in your RePEc Author Service profile, as there may be some citations waiting for confirmation.

    For technical questions regarding this item, or to correct its authors, title, abstract, bibliographic or download information, contact: arXiv administrators (email available below). General contact details of provider: http://arxiv.org/ .

    Please note that corrections may take a couple of weeks to filter through the various RePEc services.

    IDEAS is a RePEc service. RePEc uses bibliographic data supplied by the respective publishers.