EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC 4.0

Learnable Reward Weighting in Multimodal RLHF: A Proximal Policy Optimization Framework for Safe and Helpful Dialogue Alignment

Jyotsna Shastry, Shweta Agrawal

International Journal of Research and Review in Applied Science, Humanities, and Technology · 01.06.2026 · 3 (2) · 191-197

DOI: 10.71143/216rys72 ↗

Kurz gefasst

Fachartikel in der Zeitschrift International Journal of Research and Review in Applied Science, Humanities, and Technology, online erschienen am 01.06.2026. Verfasst von Jyotsna Shastry und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

In multimodal large language models, one of the major challenges is aligning these models with human values in a correct way. Traditional supervised fine-tuning methods often produce outputs that contain demographic bias, factual inaccuracies, or harmful content, as the token-prediction objective does not directly penalise these issues. To address these issues, Reinforcement Learning from Human Feedback (RLHF) has been extended to multimodal settings. However, this task is difficult because different modalities have different representations, and there is also a trade-off between helpfulness and safety. In this research, we propose a Proximal Policy Optimisation (PPO)-based RLHF framework to address these challenges. The framework introduces several innovative techniques to improve alignment and safety in multimodal environments. Evaluation on the Wizard of Wikipedia benchmark shows that, compared with the supervised fine-tuning baseline, the proposed framework achieves a 36.08% improvement in helpfulness accuracy and a 48.55% reduction in the harmfulness rate.

Zitieren

APA

Shastry, J., & Agrawal, S. (2026). Learnable Reward Weighting in Multimodal RLHF: A Proximal Policy Optimization Framework for Safe and Helpful Dialogue Alignment. International Journal of Research and Review in Applied Science, Humanities, and Technology, 3(2), 191-197. https://doi.org/10.71143/216rys72

Vancouver

Shastry J, Agrawal S. Learnable Reward Weighting in Multimodal RLHF: A Proximal Policy Optimization Framework for Safe and Helpful Dialogue Alignment. International Journal of Research and Review in Applied Science, Humanities, and Technology. 2026;3(2):191-197. doi:10.71143/216rys72

BibTeX
@article{Shastry2026,
  author = {Shastry, Jyotsna and Agrawal, Shweta},
  title = {Learnable Reward Weighting in Multimodal RLHF: A Proximal Policy Optimization Framework for Safe and Helpful Dialogue Alignment},
  journal = {International Journal of Research and Review in Applied Science, Humanities, and Technology},
  year = {2026},
  volume = {3},
  number = {2},
  pages = {191--197},
  doi = {10.71143/216rys72}
}
Kostenlos starten →