Large language models (LLMs) are susceptible to a type of attack known as jailbreaking, which misleads LLMs to output harmful contents. Although there are diverse jailbreak attack strategies, there is no unified understanding on why some methods succeed and others fail. This paper explores the behavior of harmful and harmless prompts in the LLM’s representation space to investigate the intrinsic properties of successful jailbreak attacks. We hypothesize that successful attacks share some similar properties: They are effective in moving the representation of the harmful prompt towards the direction to the harmless prompts. We leverage hidden representations into the objective of existing jailbreak attacks to move the attacks along the acceptance direction, and conduct experiments to validate the above hypothesis using the proposed objective. We hope this study provides new insights into understanding how LLMs understand harmfulness information.

Large Language Models (LLMs) are capable of producing content that perpetuates stereotypes, discrimination, and toxicity.The recently proposed moral self-correction is a computationally efficient method for reducing harmful content in the responses of LLMs. However, the process of how injecting self-correction instructions can modify the behavior of LLMs remains under-explored. In this paper, we explore the effectiveness of moral self-correction by answering three research questions: (1) In what scenarios does moral self-correction work? (2) What are the internal mechanisms of LLMs, e.g., hidden states, that are influenced by moral self-correction instructions? (3) Is intrinsic moral self-correction actually superficial in terms of reduced immorality in hidden states? We argue that self-correction can help LLMs find a shortcut to more morally correct output, rather than truly reducing the immorality stored in hidden states.Through empirical investigation with tasks of language generation and multi-choice question answering, we conclude: (i) LLMs exhibit good performance across both tasks, and self-correction instructions are particularly beneficial when the correct answer is already top-ranked; (ii) The morality levels in intermediate hidden states are strong indicators as to whether one instruction would be more effective than another; (iii) Based on our analysis of intermediate hidden states and task case studies of self-correction behaviors, we are first to propose the hypothesis that intrinsic moral self-correction is in fact superficial.

Large language models (LLMs) have show their remarkable ability in various natural language tasks. However, there are concerns that LLMs are possible to be used improperly or even illegally. To prevent the malicious usage of LLMs, detecting LLM-generated text becomes crucial in the deployment of LLM applications. Watermarking is an effective strategy to detect the LLM-generated content by encoding a pre-defined secret watermark to facilitate the detection process. However, the majority of existing watermark methods leverage the simple hashes of precedent tokens to partition vocabulary. Such watermarks can be easily eliminated by paraphrase and, correspondingly, the detection effectiveness will be greatly compromised. Thus, to enhance the robustness against paraphrase, we propose a semantics-based watermark framework, SemaMark. It leverages the semantics as an alternative to simple hashes of tokens since the semantic meaning of the sentences will be likely preserved under paraphrase and the watermark can remain robust. Comprehensive experiments are conducted to demonstrate the effectiveness and robustness of SemaMark under different paraphrases.

Retrieval-augmented generation (RAG) is a powerful technique to facilitate language model generation with proprietary and private data, where data privacy is a pivotal concern. Whereas extensive research has demonstrated the privacy risks of large language models (LLMs), the RAG technique could potentially reshape the inherent behaviors of LLM generation, posing new privacy issues that are currently under-explored. To this end, we conduct extensive empirical studies with novel attack methods, which demonstrate the vulnerability of RAG systems on leaking the private retrieval database. Despite the new risks brought by RAG on the retrieval data, we further discover that RAG can be used to mitigate the old risks, i.e., the leakage of the LLMs’ training data. In general, we reveal many new insights in this paper for privacy protection of retrieval-augmented LLMs, which could benefit both LLMs and RAG systems builders.

With the recent advancement of Large Language Models (LLMs), efforts have been made to leverage LLMs in crucial social science study methods, including predicting human features of social life such as presidential voting. Existing works suggest that LLMs are capable of generating human-like responses. Nevertheless, it is unclear how well LLMs work and where the plausible predictions derive from. This paper critically examines the performance of LLMs as social predictors, pointing out the source of correct predictions and limitations. Based on the notion of mutability that classifies social features, we design three realistic settings and a novel social prediction task, where the LLMs make predictions with input features of the same mutability and accessibility with the response feature. We find that the promising performance achieved by previous studies is because of input shortcut features to the response, which are hard to capture in reality; the performance degrades dramatically to near-random after removing the shortcuts. With the comprehensive investigations on various LLMs, we reveal that LLMs struggle to work as expected on social prediction when given ordinarily available input features without shortcuts. We further investigate possible reasons for this phenomenon and suggest potential ways to enhance LLMs for social prediction.

Large language models, such as ChatGPT, achieve amazing performance on various language processing tasks. However, they can also be exploited for improper purposes such as plagiarism or misinformation dissemination. Thus, there is an urgent need to detect the texts generated by LLMs. One type of most studied methods trains classification models to distinguish LLM texts from human texts. However, existing studies demonstrate the trained models may suffer from distribution shifts (during test), i.e., they are ineffective to predict the generated texts from unseen language tasks or topics which are not collected during training. In this work, we focus on ChatGPT as a representative model, and we conduct a comprehensive investigation on these methods’ generalization behaviors under distribution shift caused by a wide range of factors, including prompts, text lengths, topics, and language tasks. To achieve this goal, we first collect a new dataset with human and ChatGPT texts, and then we conduct extensive studies on the collected dataset. Our studies unveil insightful findings that provide guidance for future methodologies and data collection strategies for LLM detection.

Large language models (LLMs) have shown great capabilities in various tasks but also exhibited memorization of training data, raising tremendous privacy and copyright concerns. While prior works have studied memorization during pre-training, the exploration of memorization during fine-tuning is rather limited. Compared to pre-training, fine-tuning typically involves more sensitive data and diverse objectives, thus may bring distinct privacy risks and unique memorization behaviors. In this work, we conduct the first comprehensive analysis to explore language models’ (LMs) memorization during fine-tuning across tasks. Our studies with open-sourced and our own fine-tuned LMs across various tasks indicate that memorization presents a strong disparity among different fine-tuning tasks. We provide an intuitive explanation of this task disparity via sparse coding theory and unveil a strong correlation between memorization and attention score distribution.

Knowledge graphs (KGs) facilitate a wide variety of applications. Despite great efforts in creation and maintenance, even the largest KGs are far from complete. Hence, KG completion (KGC) has become one of the most crucial tasks for KG research. Recently, considerable literature in this space has centered around the use of Message Passing (Graph) Neural Networks (MPNNs), to learn powerful embeddings. The success of these methods is naturally attributed to the use of MPNNs over simpler multi-layer perceptron (MLP) models, given their additional message passing (MP) component. In this work, we find that surprisingly, simple MLP models are able to achieve comparable performance to MPNNs, suggesting that MP may not be as crucial as previously believed. With further exploration, we show careful scoring function and loss function design has a much stronger influence on KGC model performance. This suggests a conflation of scoring function design, loss function design, and MP in prior work, with promising insights regarding the scalability of state-of-the-art KGC methods today, as well as careful attention to more suitable MP designs for KGC tasks tomorrow.

Knowledge graph completion (KGC) aims to predict unseen edges in knowledge graphs (KGs), resulting in the discovery of new facts. A new class of methods have been proposed to tackle this problem by aggregating path information. These methods have shown tremendous ability in the task of KGC. However they are plagued by efficiency issues. Though there are a few recent attempts to address this through learnable path pruning, they often sacrifice the performance to gain efficiency. In this work, we identify two intrinsic limitations of these methods that affect the efficiency and representation quality. To address the limitations, we introduce a new method, TAGNet, which is able to efficiently propagate information. This is achieved by only aggregating paths in a fixed window for each source-target pair. We demonstrate that the complexity of TAGNet is independent of the number of layers. Extensive experiments demonstrate that TAGNet can cut down on the number of propagated messages by as much as 90% while achieving competitive performance on multiple KG datasets.

Crowdsourcing has emerged as a popular approach for collecting annotated data to train supervised machine learning models. However, annotator bias can lead to defective annotations. Though there are a few works investigating individual annotator bias, the group effects in annotators are largely overlooked. In this work, we reveal that annotators within the same demographic group tend to show consistent group bias in annotation tasks and thus we conduct an initial study on annotator group bias. We first empirically verify the existence of annotator group bias in various real-world crowdsourcing datasets. Then, we develop a novel probabilistic graphical framework GroupAnno to capture annotator group bias with an extended Expectation Maximization (EM) algorithm. We conduct experiments on both synthetic and real-world datasets. Experimental results demonstrate the effectiveness of our model in modeling annotator group bias in label aggregation and model learning over competitive baselines.

Recent studies show that NLP models trained on standard English texts tend to produce biased outcomes against underrepresented English varieties. In this work, we conduct a pioneering study of the English variety use of African American English (AAE) in NLI task. First, we propose CodeSwitch, a greedy unidirectional morphosyntactically-informed rule-based translation method for data augmentation. Next, we use CodeSwitch to present a preliminary study to determine if demographic language features do in fact influence models to produce false predictions. Then, we conduct experiments on two popular datasets and propose two simple, yet effective and generalizable debiasing methods. Our findings show that NLI models (e.g. BERT) trained under our proposed frameworks outperform traditional large language models while maintaining or even improving the prediction performance. In addition, we intend to release CodeSwitch, in hopes of promoting dialectal language diversity in training data to both reduce the discriminatory societal impacts and improve model robustness of downstream NLP tasks.

The automatic feedback of school assignments is an important application of AI in education. In this work, we focus on the task of personalized multimodal feedback generation, which aims to generate personalized feedback for teachers to evaluate students’ assignments involving multimodal inputs such as images, audios, and texts. This task involves the representation and fusion of multimodal information and natural language generation, which presents the challenges from three aspects: (1) how to encode and integrate multimodal inputs; (2) how to generate feedback specific to each modality; and (3) how to fulfill personalized feedback generation. In this paper, we propose a novel Personalized Multimodal Feedback Generation Network (PMFGN) armed with a modality gate mechanism and a personalized bias mechanism to address these challenges. Extensive experiments on real-world K-12 education data show that our model significantly outperforms baselines by generating more accurate and diverse feedback. In addition, detailed ablation experiments are conducted to deepen our understanding of the proposed framework.

Recently there are increasing concerns about the fairness of Artificial Intelligence (AI) in real-world applications such as computer vision and recommendations. For example, recognition algorithms in computer vision are unfair to black people such as poorly detecting their faces and inappropriately identifying them as “gorillas”. As one crucial application of AI, dialogue systems have been extensively applied in our society. They are usually built with real human conversational data; thus they could inherit some fairness issues which are held in the real world. However, the fairness of dialogue systems has not been well investigated. In this paper, we perform a pioneering study about the fairness issues in dialogue systems. In particular, we construct a benchmark dataset and propose quantitative measures to understand fairness in dialogue models. Our studies demonstrate that popular dialogue models show significant prejudice towards different genders and races. Besides, to mitigate the bias in dialogue systems, we propose two simple but effective debiasing methods. Experiments show that our methods can reduce the bias in dialogue systems significantly. The dataset and the implementation are released to foster fairness research in dialogue systems.

Dialogue systems play an increasingly important role in various aspects of our daily life. It is evident from recent research that dialogue systems trained on human conversation data are biased. In particular, they can produce responses that reflect people’s gender prejudice. Many debiasing methods have been developed for various NLP tasks, such as word embedding. However, they are not directly applicable to dialogue systems because they are likely to force dialogue models to generate similar responses for different genders. This greatly degrades the diversity of the generated responses and immensely hurts the performance of the dialogue models. In this paper, we propose a novel adversarial learning framework Debiased-Chat to train dialogue models free from gender bias while keeping their performance. Extensive experiments on two real-world conversation datasets show that our framework significantly reduces gender bias in dialogue models while maintaining the response quality.

On the one hand, nowadays, fake news articles are easily propagated through various online media platforms and have become a grand threat to the trustworthiness of information. On the other hand, our understanding of the language of fake news is still minimal. Incorporating hierarchical discourse-level structure of fake and real news articles is one crucial step toward a better understanding of how these articles are structured. Nevertheless, this has rarely been investigated in the fake news detection domain and faces tremendous challenges. First, existing methods for capturing discourse-level structure rely on annotated corpora which are not available for fake news datasets. Second, how to extract out useful information from such discovered structures is another challenge. To address these challenges, we propose Hierarchical Discourse-level Structure for Fake news detection. HDSF learns and constructs a discourse-level structure for fake/real news articles in an automated and data-driven manner. Moreover, we identify insightful structure-related properties, which can explain the discovered structures and boost our understating of fake news. Conducted experiments show the effectiveness of the proposed approach. Further structural analysis suggests that real and fake news present substantial differences in the hierarchical discourse-level structures.

Fake news spreading through media outlets poses a real threat to the trustworthiness of information and detecting fake news has attracted increasing attention in recent years. Fake news is typically written intentionally to mislead readers, which determines that fake news detection merely based on news content is tremendously challenging. Meanwhile, fake news could contain true evidence to mock true news and presents different degrees of fakeness, which further exacerbates the detection difficulty. On the other hand, the spread of fake news produces various types of data from different perspectives. These multiple sources provide rich contextual information about fake news and offer unprecedented opportunities for advanced fake news detection. In this paper, we study fake news detection with different degrees of fakeness by integrating multiple sources. In particular, we introduce approaches to combine information from multiple sources and to discriminate between different degrees of fakeness, and propose a Multi-source Multi-class Fake news Detection framework MMFD, which combines automated feature extraction, multi-source fusion and automated degrees of fakeness detection into a coherent and interpretable model. Experimental results on the real-world data demonstrate the effectiveness of the proposed framework and extensive experiments are further conducted to understand the working of the proposed framework.