Reinforcement Learning from Human Feedback (RLHF) has become an essential technique for enhancing pre-trained large language models (LLMs) to generate responses that align with human preferences and societal values. Although RLHF has shown promise, the training of reward models (RMs) still faces the challenge of reward hacking, motivating recent works to prevent RMs from finding shortcuts that bypass the intended optimization objectives by identifying simplistic patterns such as response length. Besides the issue of length bias, our work firstly reveals that prompt-template bias learned by RMs can also cause reward hacking when dealing with some marginal samples, resulting in LLMs preferring to generate responses in a specific format after RLHF fine-tuning, regardless of the format requested in the prompt. To this end, we propose a low-cost but effective method, namely Prompt Bias Calibration (PBC), to estimate the prompt-template bias term during reward modeling, which can be utilized to calibrate reward scores in the following RL fine-tuning process. Then, we show that our PBC method can be flexibly combined with existing algorithms of removing length bias, leading to a further improvement in the aspect of enhancing the quality of generated responses.

Natural language processing (NLP) often faces the problem of data diversity such as different domains, themes, styles, and so on. Therefore, a single language model (LM) is insufficient to learn all knowledge from diverse samples. To solve this problem, we firstly propose an autoencoding topic model with a mixture prior (mATM) to perform clustering for the data, where the clusters defined in semantic space describes the data diversity. Having obtained the clustering assignment for each sample, we develop the ensemble LM (EnsLM) with the technique of weight modulation. Specifically, EnsLM contains a backbone that is adjusted by a few modulated weights to fit for different sample clusters. As a result, the backbone learns the shared knowledge among all clusters while modulated weights extract the cluster-specific features. EnsLM can be trained jointly with mATM with a flexible LM backbone. We evaluate the effectiveness of both mATM and EnsLM on various tasks.

Abstractive document summarization is a comprehensive task including document understanding and summary generation, in which area Transformer-based models have achieved the state-of-the-art performance. Compared with Transformers, topic models are better at learning explicit document semantics, and hence could be integrated into Transformers to further boost their performance. To this end, we rearrange and explore the semantics learned by a topic model, and then propose a topic assistant (TA) including three modules. TA is compatible with various Transformer-based models and user-friendly since i) TA is a plug-and-play model that does not break any structure of the original Transformer network, making users easily fine-tune Transformer+TA based on a well pre-trained model; ii) TA only introduces a small number of extra parameters. Experimental results on three datasets demonstrate that TA is able to improve the performance of several Transformer-based models.