Pengfei Xu

Jian Li, Haojing Huang, Yujia Zhang et al.

Recently, there has been significant interest in replacing the reward model in Reinforcement Learning with Human Feedback (RLHF) methods for Large Language Models (LLMs), such as Direct Preference Optimization (DPO) and its variants. These approaches commonly use a binary cross-entropy mechanism on pairwise samples, i.e., minimizing and maximizing the loss based on preferred or dis-preferred responses, respectively. However, while this training strategy omits the reward model, it also overlooks the varying preference degrees within different responses. We hypothesize that this is a key factor hindering LLMs from sufficiently understanding human preferences. To address this problem, we propose a novel Self-supervised Preference Optimization (SPO) framework, which constructs a self-supervised preference degree loss combined with the alignment loss, thereby helping LLMs improve their ability to understand the degree of preference. Extensive experiments are conducted on two widely used datasets of different tasks. The results demonstrate that SPO can be seamlessly integrated with existing preference optimization methods and significantly boost their performance to achieve state-of-the-art performance. We also conduct detailed analyses to offer comprehensive insights into SPO, which verifies its effectiveness. The code is available at https://github.com/lijian16/SPO.

Sicheng Zhao, Hui Chen, Hu Huang et al.

In many practical applications, it is often difficult and expensive to obtain large-scale labeled data to train state-of-the-art deep neural networks. Therefore, transferring the learned knowledge from a separate, labeled source domain to an unlabeled or sparsely labeled target domain becomes an appealing alternative. However, direct transfer often results in significant performance decay due to domain shift. Domain adaptation (DA) aims to address this problem by aligning the distributions between the source and target domains. Multi-source domain adaptation (MDA) is a powerful and practical extension in which the labeled data may be collected from multiple sources with different distributions. In this survey, we first define various MDA strategies. Then we systematically summarize and compare modern MDA methods in the deep learning era from different perspectives, followed by commonly used datasets and a brief benchmark. Finally, we discuss future research directions for MDA that are worth investigating.

Jiankun Zhu, Sicheng Zhao, Jing Jiang et al.

Visual emotion recognition (VER), which aims at understanding humans' emotional reactions toward different visual stimuli, has attracted increasing attention. Given the subjective and ambiguous characteristics of emotion, annotating a reliable large-scale dataset is hard. For reducing reliance on data labeling, domain adaptation offers an alternative solution by adapting models trained on labeled source data to unlabeled target data. Conventional domain adaptation methods require access to source data. However, due to privacy concerns, source emotional data may be inaccessible. To address this issue, we propose an unexplored task: source-free domain adaptation (SFDA) for VER, which does not have access to source data during the adaptation process. To achieve this, we propose a novel framework termed Bridge then Begin Anew (BBA), which consists of two steps: domain-bridged model generation (DMG) and target-related model adaptation (TMA). First, the DMG bridges cross-domain gaps by generating an intermediate model, avoiding direct alignment between two VER datasets with significant differences. Then, the TMA begins training the target model anew to fit the target structure, avoiding the influence of source-specific knowledge. Extensive experiments are conducted on six SFDA settings for VER. The results demonstrate the effectiveness of BBA, which achieves remarkable performance gains compared with state-of-the-art SFDA methods and outperforms representative unsupervised domain adaptation approaches.

Jian Li, Shenglin Yin, Yujia Zhang et al.

Direct Preference Optimization (DPO) is a widely used reinforcement learning from human feedback (RLHF) method across various domains. Recent research has increasingly focused on the role of token importance in improving DPO effectiveness. It is observed that identical or semantically similar content (defined as ambiguous content) frequently appears within the preference pairs. We hypothesize that the presence of ambiguous content during DPO training may introduce ambiguity, thereby limiting further improvements in alignment. Through mathematical analysis and proof-of-concept experiments, we reveal that ambiguous content may potentially introduce ambiguities, thereby degrading performance. To address this issue, we introduce Ambiguity Awareness Optimization (AAO), a simple yet effective approach that automatically re-weights ambiguous content to reduce ambiguities by calculating semantic similarity from preference pairs. Through extensive experiments, we demonstrate that AAO consistently and significantly surpasses state-of-the-art approaches in performance, without markedly increasing response length, across multiple model scales and widely adopted benchmark datasets, including AlpacaEval 2, MT-Bench, and Arena-Hard. Specifically, AAO outperforms DPO by up to 8.9 points on AlpacaEval 2 and achieves an improvement of by up to 15.0 points on Arena-Hard.

Xiuwen Wu, Hao Wang, Zhiang Yan et al.

Effective and safe human-machine collaboration requires the regulated and meaningful exchange of emotions between humans and artificial intelligence (AI). Current AI systems based on large language models (LLMs) can provide feedback that makes people feel heard. Yet it remains unclear whether LLMs represent emotion in language as humans do, or whether and how the emotional tone of their output can be controlled. We assess human-AI emotional alignment across linguistic-cultural groups and model-families, using interpretable LLM features translated from concept-sets for over twenty nuanced emotion categories (including six basic emotions). Our analyses reveal that LLM-derived emotion spaces are structurally congruent with human perception, underpinned by the fundamental affective dimensions of valence and arousal. Furthermore, these emotion-related features also accurately predict large-scale behavioural data on word ratings along these two core dimensions, reflecting both universal and language-specific patterns. Finally, by leveraging steering vectors derived solely from human-centric emotion concepts, we show that model expressions can be stably and naturally modulated across distinct emotion categories, which provides causal evidence that human emotion concepts can be used to systematically induce LLMs to produce corresponding affective states when conveying content. These findings suggest AI not only shares emotional representations with humans but its affective outputs can be precisely guided using psychologically grounded emotion concepts.