Yichen Ouyang, Lu Wang, Fangkai Yang, Pu Zhao, Chenghua Huang, Jianfeng Liu, Bochen Pang, Yaming Yang, Yuefeng Zhan, Hao Sun, Qingwei Lin, Saravan Rajmohan, Weiwei Deng, Dongmei Zhang, Feng Sun. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Reinforcement Learning from Human Feedback (RLHF) is a crucial technique for aligning language models with human preferences and is a key factor in the success of modern conversational models like GPT-4, ChatGPT, and Llama 2. A significant challenge in employing RLHF lies in training a reliable RM, which relies on high-quality labels. Typically, these labels are provided by human experts or a stronger AI, both of which can be costly and introduce bias that may affect the language model's responses. As models improve, human input may become less effective in enhancing their performance. This paper explores the potential of using the RM itself to generate additional training data for a more robust RM. Our experiments demonstrate that reinforcement learning from self-feedback outperforms baseline approaches. We conducted extensive experiments with our approach on multiple datasets, such as HH-RLHF and UltraFeedback, and models including Mistral and Llama 3, comparing it against various baselines. Our results indicate that, even with a limited amount of human-labeled data, learning from self-feedback can robustly enhance the performance of the RM, thereby improving the capabilities of large language models.

Semi-Supervised Relation Extraction aims at learning well-performed RE models with limited labeled and large-scale unlabeled data. Existing methods mainly suffer from semantic drift and insufficient supervision, which severely limit the performance. To address these problems, recent work tends to design dual modules to work cooperatively for mutual enhancement. However, the consensus of two modules greatly restricts the model from exploring diverse relation expressions in unlabeled set, which hinders the performance as well as model generalization. To tackle this problem, in this paper, we propose a novel competition-based method AdvSRE. We set up a challenging minimax game on unlabeled data between two modules, Generator and Discriminator, and assign them with conflicting objectives. During the competition game, one module may find any possible chance to beat the other, which develops two modules' abilities until relation expressions cannot be further explored. To exploit label information, Discriminator is further asked to predict specific relation for each sentence. Experiment results on two benchmarks show new state-of-the-art performance over baselines, demonstrating the effectiveness of proposed AdvSRE.