Despite the efficacy of Direct Preference Optimization (DPO) in aligning Large Language Models (LLMs), reward hacking remains a pivotal challenge. This issue emerges when LLMs excessively reduce the probability of rejected completions to achieve high rewards, without genuinely meeting their intended goals. As a result, this leads to overly lengthy generation lacking diversity, as well as catastrophic forgetting of knowledge. We investigate the underlying reason behind this issue, which is representation redundancy caused by neuron collapse in the parameter space. Hence, we propose a novel Weights-Rotated Preference Optimization (RoPO) algorithm, which implicitly constrains the output layer logits with the KL divergence inherited from DPO and explicitly constrains the intermediate hidden states by fine-tuning on a multi-granularity orthogonal matrix. This design prevents the policy model from deviating too far from the reference model, thereby retaining the knowledge and expressive capabilities acquired during pre-training and SFT stages. Our RoPO achieves up to a 3.27-point improvement on AlpacaEval 2, and surpasses the best baseline by 6.2 to 7.5 points on MT-Bench with merely 0.015% of the trainable parameters, demonstrating its effectiveness in alleviating the reward hacking problem of DPO.

Recent years have witnessed the rise and success of pre-training techniques in visually-rich document understanding. However, most existing methods lack the systematic mining and utilization of layout-centered knowledge, leading to sub-optimal performances. In this paper, we propose ERNIE-Layout, a novel document pre-training solution with layout knowledge enhancement in the whole workflow, to learn better representations that combine the features from text, layout, and image. Specifically, we first rearrange input sequences in the serialization stage, and then present a correlative pre-training task, reading order prediction, to learn the proper reading order of documents. To improve the layout awareness of the model, we integrate a spatial-aware disentangled attention into the multi-modal transformer and a replaced regions prediction task into the pre-training phase. Experimental results show that ERNIE-Layout achieves superior performance on various downstream tasks, setting new state-of-the-art on key information extraction, document image classification, and document question answering datasets. The code and models are publicly available at PaddleNLP.

This paper describes our system participated in Task 6 of SemEval-2021: the task focuses on multimodal propaganda technique classification and it aims to classify given image and text into 22 classes. In this paper, we propose to use transformer based architecture to fuse the clues from both image and text. We explore two branches of techniques including fine-tuning the text pretrained transformer with extended visual features, and fine-tuning the multimodal pretrained transformers. For the visual features, we have tested both grid features based on ResNet and salient region features from pretrained object detector. Among the pretrained multimodal transformers, we choose ERNIE-ViL, a two-steam cross-attended transformers pretrained on large scale image-caption aligned data. Fine-tuing ERNIE-ViL for our task produce a better performance due to general joint multimodal representation for text and image learned by ERNIE-ViL. Besides, as the distribution of the classification labels is very unbalanced, we also make a further attempt on the loss function and the experiment result shows that focal loss would perform better than cross entropy loss. Last we have won first for subtask C in the final competition.