Pairwise Preference Reward and Group-Based Diversity Enhancement for Superior Open-Ended Generation
The paper introduces a new reinforcement learning method called Pairwise Preference Reward and Group-based Diversity Enhancement (PPR-GDE) for open-ended generation tasks. This method aims to address challenges such as diversity collapse and high computational costs associated with traditional reward models. Experimental results indicate that PPR-GDE achieves better alignment quality and expressive diversity compared to existing reinforcement learning baselines.
- ▪PPR-GDE does not require scalar rewards and incorporates group-level diversity into the reward signal.
- ▪The method preserves the comparative structure of subjective evaluation through a pairwise preference reward.
- ▪Experiments show that PPR-GDE outperforms strong RL baselines in terms of alignment quality and diversity.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.18191 (cs) [Submitted on 18 May 2026] Title:Pairwise Preference Reward and Group-Based Diversity Enhancement for Superior Open-Ended Generation Authors:Guining Cao, Jiaxin Peng, Chu Zeng, Yu Zhao, Shuangyong Song, Yongxiang View a PDF of the paper titled Pairwise Preference Reward and Group-Based Diversity Enhancement for Superior Open-Ended Generation, by Guining Cao and 5 other authors View PDF HTML (experimental) Abstract:Current reinforcement learning(RL) methods are broadly applicable and powerful in verifiable settings where scalar rewards can be provided.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
