Large language models (LLMs) have recently experienced remarkable growth in both their capabilities (OpenAI, 2023; Gemini Team et al., 2023; Touvron et al., 2023) and their applications in various fields (NLLB Team et al., 2022; Thirunavukarasu et al., 2023; Schick et al., 2023; Bubeck et al., 2023). As LLMs become increasingly complex and are deployed in safety-critical environments (Singhal et al., 2022; Li et al., 2023; Maddela et al., 2023), it is essential to thoroughly understand their robustness to different inputs. Indeed, the susceptibility of LLMs to user inputs and adversarial prompts -- prompts crafted to mislead the model or exploit its weaknesses, potentially leading to unsafe, biased, or incorrect outputs -- poses a significant challenge (Perez et al., 2022; Wei et al., 2023; Zou et al., 2023). Identifying these vulnerabilities and subsequently mitigating such risks is therefore vital to ensure the safe and reliable operation of LLMs in the real world. Current methods for identifying adversarial prompts aimed at "attacking" LLMs and eliciting undesirable outputs are limited by several factors.

Adversarial misuse, particularly through `jailbreaking' that circumvents a model's safety and ethical protocols, poses a significant challenge for Large Language Models (LLMs). This paper delves into the mechanisms behind such successful attacks, introducing a hypothesis for the safety mechanism of aligned LLMs: intent security recognition followed by response generation. Grounded in this hypothesis, we propose CodeChameleon, a novel jailbreak framework based on personalized encryption tactics. To elude the intent security recognition phase, we reformulate tasks into a code completion format, enabling users to encrypt queries using personalized encryption functions. To guarantee response generation functionality, we embed a decryption function within the instructions, which allows the LLM to decrypt and execute the encrypted queries successfully. We conduct extensive experiments on 7 LLMs, achieving state-of-the-art average Attack Success Rate (ASR). Remarkably, our method achieves an 86.6\% ASR on GPT-4-1106.

Machine learning (ML) is widely used to moderate online content. Despite its scalability relative to human moderation, the use of ML introduces unique challenges to content moderation. One such challenge is predictive multiplicity: multiple competing models for content classification may perform equally well on average, yet assign conflicting predictions to the same content. This multiplicity can result from seemingly innocuous choices during model development, such as random seed selection for parameter initialization. We experimentally demonstrate how content moderation tools can arbitrarily classify samples as toxic, leading to arbitrary restrictions on speech. We discuss these findings in terms of human rights set out by the International Covenant on Civil and Political Rights (ICCPR), namely freedom of expression, non-discrimination, and procedural justice. We analyze (i) the extent of predictive multiplicity among state-of-the-art LLMs used for detecting toxic content; (ii) the disparate impact of this arbitrariness across social groups; and (iii) how model multiplicity compares to unambiguous human classifications. Our findings indicate that the up-scaled algorithmic moderation risks legitimizing an algorithmic leviathan, where an algorithm disproportionately manages human rights. To mitigate such risks, our study underscores the need to identify and increase the transparency of arbitrariness in content moderation applications. Since algorithmic content moderation is being fueled by pressing social concerns, such as disinformation and hate speech, our discussion on harms raises concerns relevant to policy debates. Our findings also contribute to content moderation and intermediary liability laws being discussed and passed in many countries, such as the Digital Services Act in the European Union, the Online Safety Act in the United Kingdom, and the Fake News Bill in Brazil.

When completing knowledge-intensive tasks, humans sometimes need not just an answer but also a corresponding reference passage for auxiliary reading. Previous methods required obtaining pre-segmented article chunks through additional retrieval models. This paper explores leveraging the parameterized knowledge stored during the pre-training phase of large language models (LLMs) to independently recall reference passage from any starting position. We propose a two-stage framework that simulates the scenario of humans recalling easily forgotten references. Initially, the LLM is prompted to recall document title identifiers to obtain a coarse-grained document set. Then, based on the acquired coarse-grained document set, it recalls fine-grained passage. In the two-stage recall process, we use constrained decoding to ensure that content outside of the stored documents is not generated. To increase speed, we only recall a short prefix in the second stage, then locate its position to retrieve a complete passage. Experiments on KILT knowledge-sensitive tasks have verified that LLMs can independently recall reference passage location in various task forms, and the obtained reference significantly assist downstream tasks.

The safety of Large Language Models (LLMs) has gained increasing attention in recent years, but there still lacks a comprehensive approach for detecting safety issues within LLMs' responses in an aligned, customizable and explainable manner. In this paper, we propose ShieldLM, an LLM-based safety detector, which aligns with general human safety standards, supports customizable detection rules, and provides explanations for its decisions. To train ShieldLM, we compile a large bilingual dataset comprising 14,387 query-response pairs, annotating the safety of responses based on various safety standards. Through extensive experiments, we demonstrate that ShieldLM surpasses strong baselines across four test sets, showcasing remarkable customizability and explainability. Besides performing well on standard detection datasets, ShieldLM has also been shown to be effective in real-world situations as a safety evaluator for advanced LLMs. We release ShieldLM at \url{https://github.com/thu-coai/ShieldLM} to support accurate and explainable safety detection under various safety standards, contributing to the ongoing efforts to enhance the safety of LLMs.

Generative AI platforms and features are permeating many aspects of work. Entrepreneurs from lean economies in particular are well positioned to outsource tasks to generative AI given limited resources. In this paper, we work to address a growing disparity in use of these technologies by building on a four-year partnership with a local entrepreneurial hub dedicated to equity in tech and entrepreneurship. Together, we co-designed an interactive workshops series aimed to onboard local entrepreneurs to generative AI platforms. Alongside four community-driven and iterative workshops with entrepreneurs across five months, we conducted interviews with 15 local entrepreneurs and community providers. We detail the importance of communal and supportive exposure to generative AI tools for local entrepreneurs, scaffolding actionable use (and supporting non-use), demystifying generative AI technologies by Figure 1: We designed an introductory generative AI workshop emphasizing entrepreneurial power, while simultaneously deconstructing series with entrepreneurs and tech providers which centered the veneer of simplicity to address the many operational communal experience, supportive exposure, tangible skills needed for successful application.

Generative AI image models may inadvertently generate problematic representations of people. Past research has noted that millions of users engage daily across the world with these models and that the models, including through problematic representations of people, have the potential to compound and accelerate real-world discrimination and other harms (Bianchi et al, 2023). In this paper, we focus on addressing the generation of problematic associations between demographic groups and semantic concepts that may reflect and reinforce negative narratives embedded in social data. Building on sociological literature (Blumer, 1958) and mapping representations to model behaviors, we have developed a taxonomy to study problematic associations in image generation models. We explore the effectiveness of fine tuning at the model level as a method to address these associations, identifying a potential reduction in visual quality as a limitation of traditional fine tuning. We also propose a new methodology with twice-human-in-the-loop (T-HITL) that promises improvements in both reducing problematic associations and also maintaining visual quality. We demonstrate the effectiveness of T-HITL by providing evidence of three problematic associations addressed by T-HITL at the model level. Our contributions to scholarship are two-fold. By defining problematic associations in the context of machine learning models and generative AI, we introduce a conceptual and technical taxonomy for addressing some of these associations. Finally, we provide a method, T-HITL, that addresses these associations and simultaneously maintains visual quality of image model generations. This mitigation need not be a tradeoff, but rather an enhancement.

Recent advances in pre-trained language modeling have facilitated significant progress across various natural language processing (NLP) tasks. Word masking during model training constitutes a pivotal component of language modeling in architectures like BERT. However, the prevalent method of word masking relies on random selection, potentially disregarding domain-specific linguistic attributes. In this article, we introduce an innovative masking approach leveraging genre and topicality information to tailor language models to specialized domains. Our method incorporates a ranking process that prioritizes words based on their significance, subsequently guiding the masking procedure. Experiments conducted using continual pre-training within the legal domain have underscored the efficacy of our approach on the LegalGLUE benchmark in the English language. Pre-trained language models and code are freely available for use.

The assessment of explainability in Legal Judgement Prediction (LJP) systems is of paramount importance in building trustworthy and transparent systems, particularly considering the reliance of these systems on factors that may lack legal relevance or involve sensitive attributes. This study delves into the realm of explainability and fairness in LJP models, utilizing Swiss Judgement Prediction (SJP), the only available multilingual LJP dataset. We curate a comprehensive collection of rationales that `support' and `oppose' judgement from legal experts for 108 cases in German, French, and Italian. By employing an occlusion-based explainability approach, we evaluate the explainability performance of state-of-the-art monolingual and multilingual BERT-based LJP models, as well as models developed with techniques such as data augmentation and cross-lingual transfer, which demonstrated prediction performance improvement. Notably, our findings reveal that improved prediction performance does not necessarily correspond to enhanced explainability performance, underscoring the significance of evaluating models from an explainability perspective. Additionally, we introduce a novel evaluation framework, Lower Court Insertion (LCI), which allows us to quantify the influence of lower court information on model predictions, exposing current models' biases.

I'll tell you what mine is: death. I am not really built for battle – I need five cups of coffee just to function and I have terrible allergies. My body can't even handle pollen, it's not going to do well with nuclear war. Plus, even if I was hardier – who wants to live a few extra months in a completely destroyed world? As you have probably noticed bunkers have become the ultimate status symbol among the 1%.