Because "out-of-the-box" large language models are capable of generating a great deal of objectionable content, recent work has focused on aligning these models in an attempt to prevent undesirable generation. While there has been some success at circumventing these measures -- so-called "jailbreaks" against LLMs -- these attacks have required significant human ingenuity and are brittle in practice. In this paper, we propose a simple and effective attack method that causes aligned language models to generate objectionable behaviors. Specifically, our approach finds a suffix that, when attached to a wide range of queries for an LLM to produce objectionable content, aims to maximize the probability that the model produces an affirmative response (rather than refusing to answer). However, instead of relying on manual engineering, our approach automatically produces these adversarial suffixes by a combination of greedy and gradient-based search techniques, and also improves over past automatic prompt generation methods. Surprisingly, we find that the adversarial prompts generated by our approach are quite transferable, including to black-box, publicly released LLMs. Specifically, we train an adversarial attack suffix on multiple prompts (i.e., queries asking for many different types of objectionable content), as well as multiple models (in our case, Vicuna-7B and 13B). When doing so, the resulting attack suffix is able to induce objectionable content in the public interfaces to ChatGPT, Bard, and Claude, as well as open source LLMs such as LLaMA-2-Chat, Pythia, Falcon, and others. In total, this work significantly advances the state-of-the-art in adversarial attacks against aligned language models, raising important questions about how such systems can be prevented from producing objectionable information. Code is available at github.com/llm-attacks/llm-attacks.
相關內容
Node-link diagrams are a popular method for representing graphs that capture relationships between individuals, businesses, proteins, and telecommunication endpoints. However, node-link diagrams may fail to convey insights regarding graph structures, even for moderately sized data of a few hundred nodes, due to visual clutter. We propose to apply the mapper construction -- a popular tool in topological data analysis -- to graph visualization, which provides a strong theoretical basis for summarizing the data while preserving their core structures. We develop a variation of the mapper construction targeting weighted, undirected graphs, called {\mog}, which generates homology-preserving skeletons of graphs. We further show how the adjustment of a single parameter enables multi-scale skeletonization of the input graph. We provide a software tool that enables interactive explorations of such skeletons and demonstrate the effectiveness of our method for synthetic and real-world data.
NSFW (Not Safe for Work) content, in the context of a dialogue, can have severe side effects on users in open-domain dialogue systems. However, research on detecting NSFW language, especially sexually explicit content, within a dialogue context has significantly lagged behind. To address this issue, we introduce CensorChat, a dialogue monitoring dataset aimed at NSFW dialogue detection. Leveraging knowledge distillation techniques involving GPT-4 and ChatGPT, this dataset offers a cost-effective means of constructing NSFW content detectors. The process entails collecting real-life human-machine interaction data and breaking it down into single utterances and single-turn dialogues, with the chatbot delivering the final utterance. ChatGPT is employed to annotate unlabeled data, serving as a training set. Rationale validation and test sets are constructed using ChatGPT and GPT-4 as annotators, with a self-criticism strategy for resolving discrepancies in labeling. A BERT model is fine-tuned as a text classifier on pseudo-labeled data, and its performance is assessed. The study emphasizes the importance of AI systems prioritizing user safety and well-being in digital conversations while respecting freedom of expression. The proposed approach not only advances NSFW content detection but also aligns with evolving user protection needs in AI-driven dialogues.
Natural language is an appealing medium for explaining how large language models process and store information, but evaluating the faithfulness of such explanations is challenging. To help address this, we develop two modes of evaluation for natural language explanations that claim individual neurons represent a concept in a text input. In the observational mode, we evaluate claims that a neuron $a$ activates on all and only input strings that refer to a concept picked out by the proposed explanation $E$. In the intervention mode, we construe $E$ as a claim that the neuron $a$ is a causal mediator of the concept denoted by $E$. We apply our framework to the GPT-4-generated explanations of GPT-2 XL neurons of Bills et al. (2023) and show that even the most confident explanations have high error rates and little to no causal efficacy. We close the paper by critically assessing whether natural language is a good choice for explanations and whether neurons are the best level of analysis.
In modern commercial search engines and recommendation systems, data from multiple domains is available to jointly train the multi-domain model. Traditional methods train multi-domain models in the multi-task setting, with shared parameters to learn the similarity of multiple tasks, and task-specific parameters to learn the divergence of features, labels, and sample distributions of individual tasks. With the development of large language models, LLM can extract global domain-invariant text features that serve both search and recommendation tasks. We propose a novel framework called S\&R Multi-Domain Foundation, which uses LLM to extract domain invariant features, and Aspect Gating Fusion to merge the ID feature, domain invariant text features and task-specific heterogeneous sparse features to obtain the representations of query and item. Additionally, samples from multiple search and recommendation scenarios are trained jointly with Domain Adaptive Multi-Task module to obtain the multi-domain foundation model. We apply the S\&R Multi-Domain foundation model to cold start scenarios in the pretrain-finetune manner, which achieves better performance than other SOTA transfer learning methods. The S\&R Multi-Domain Foundation model has been successfully deployed in Alipay Mobile Application's online services, such as content query recommendation and service card recommendation, etc.
The emergence of human-like abilities of AI systems for content generation in domains such as text, audio, and vision has prompted the development of classifiers to determine whether content originated from a human or a machine. Implicit in these efforts is an assumption that the generation properties of a human are different from that of the machine. In this work, we provide a framework in the language of statistical pattern recognition that quantifies the difference between the distributions of human and machine-generated content conditioned on an evaluation context. We describe current methods in the context of the framework and demonstrate how to use the framework to evaluate the progression of generative models towards human-like capabilities, among many axes of analysis.
Humans learn about objects via interaction and using multiple perceptions, such as vision, sound, and touch. While vision can provide information about an object's appearance, non-visual sensors, such as audio and haptics, can provide information about its intrinsic properties, such as weight, temperature, hardness, and the object's sound. Using tools to interact with objects can reveal additional object properties that are otherwise hidden (e.g., knives and spoons can be used to examine the properties of food, including its texture and consistency). Robots can use tools to interact with objects and gather information about their implicit properties via non-visual sensors. However, a robot's model for recognizing objects using a tool-mediated behavior does not generalize to a new tool or behavior due to differing observed data distributions. To address this challenge, we propose a framework to enable robots to transfer implicit knowledge about granular objects across different tools and behaviors. The proposed approach learns a shared latent space from multiple robots' contexts produced by respective sensory data while interacting with objects using tools. We collected a dataset using a UR5 robot that performed 5,400 interactions using 6 tools and 6 behaviors on 15 granular objects and tested our method on cross-tool and cross-behavioral transfer tasks. Our results show the less experienced target robot can benefit from the experience gained from the source robot and perform recognition on a set of novel objects. We have released the code, datasets, and additional results: //github.com/gtatiya/Tool-Knowledge-Transfer.
In recent years, large language models (LLMs) have shown remarkable capabilities at scale, particularly at generating text conditioned on a prompt. In our work, we investigate the use of LLMs to augment training data of small language models~(SLMs) with automatically generated counterfactual~(CF) instances -- i.e. minimally altered inputs -- in order to improve out-of-domain~(OOD) performance of SLMs in the extractive question answering~(QA) setup. We show that, across various LLM generators, such data augmentation consistently enhances OOD performance and improves model calibration for both confidence-based and rationale-augmented calibrator models. Furthermore, these performance improvements correlate with higher diversity of CF instances in terms of their surface form and semantic content. Finally, we show that CF augmented models which are easier to calibrate also exhibit much lower entropy when assigning importance, indicating that rationale-augmented calibrators prefer concise explanations.
Summarization is an important application of large language models (LLMs). Most previous evaluation of summarization models has focused on their performance in content selection, grammaticality and coherence. However, it is well known that LLMs reproduce and reinforce harmful social biases. This raises the question: Do these biases affect model outputs in a relatively constrained setting like summarization? To help answer this question, we first motivate and introduce a number of definitions for biased behaviours in summarization models, along with practical measures to quantify them. Since we find biases inherent to the input document can confound our analysis, we additionally propose a method to generate input documents with carefully controlled demographic attributes. This allows us to sidestep this issue, while still working with somewhat realistic input documents. Finally, we apply our measures to summaries generated by both purpose-built summarization models and general purpose chat models. We find that content selection in single document summarization seems to be largely unaffected by bias, while hallucinations exhibit evidence of biases propagating to generated summaries.
We consider the problem of referring image segmentation. Given an input image and a natural language expression, the goal is to segment the object referred by the language expression in the image. Existing works in this area treat the language expression and the input image separately in their representations. They do not sufficiently capture long-range correlations between these two modalities. In this paper, we propose a cross-modal self-attention (CMSA) module that effectively captures the long-range dependencies between linguistic and visual features. Our model can adaptively focus on informative words in the referring expression and important regions in the input image. In addition, we propose a gated multi-level fusion module to selectively integrate self-attentive cross-modal features corresponding to different levels in the image. This module controls the information flow of features at different levels. We validate the proposed approach on four evaluation datasets. Our proposed approach consistently outperforms existing state-of-the-art methods.
Deep learning has revolutionized many machine learning tasks in recent years, ranging from image classification and video processing to speech recognition and natural language understanding. The data in these tasks are typically represented in the Euclidean space. However, there is an increasing number of applications where data are generated from non-Euclidean domains and are represented as graphs with complex relationships and interdependency between objects. The complexity of graph data has imposed significant challenges on existing machine learning algorithms. Recently, many studies on extending deep learning approaches for graph data have emerged. In this survey, we provide a comprehensive overview of graph neural networks (GNNs) in data mining and machine learning fields. We propose a new taxonomy to divide the state-of-the-art graph neural networks into different categories. With a focus on graph convolutional networks, we review alternative architectures that have recently been developed; these learning paradigms include graph attention networks, graph autoencoders, graph generative networks, and graph spatial-temporal networks. We further discuss the applications of graph neural networks across various domains and summarize the open source codes and benchmarks of the existing algorithms on different learning tasks. Finally, we propose potential research directions in this fast-growing field.
小貼士
登錄享
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191