As generative large model capabilities advance, safety concerns become more pronounced in their outputs. To ensure the sustainable growth of the AI ecosystem, it's imperative to undertake a holistic evaluation and refinement of associated safety risks. This survey presents a framework for safety research pertaining to large models, delineating the landscape of safety risks as well as safety evaluation and improvement methods. We begin by introducing safety issues of wide concern, then delve into safety evaluation methods for large models, encompassing preference-based testing, adversarial attack approaches, issues detection, and other advanced evaluation methods. Additionally, we explore the strategies for enhancing large model safety from training to deployment, highlighting cutting-edge safety approaches for each stage in building large models. Finally, we discuss the core challenges in advancing towards more responsible AI, including the interpretability of safety mechanisms, ongoing safety issues, and robustness against malicious attacks. Through this survey, we aim to provide clear technical guidance for safety researchers and encourage further study on the safety of large models.

Transformer-based language models have been shown to be highly effective for several NLP tasks. In this paper, we consider three transformer models, BERT, RoBERTa, and XLNet, in both small and large versions, and investigate how faithful their representations are with respect to the semantic content of texts. We formalize a notion of semantic faithfulness, in which the semantic content of a text should causally figure in a model's inferences in question answering. We then test this notion by observing a model's behavior on answering questions about a story after performing two novel semantic interventions: deletion intervention and negation intervention. While transformer models achieve high performance on standard question answering tasks, we show that they fail to be semantically faithful once we perform these interventions for a significant number of cases (~50% for deletion intervention, and ~20% drop in accuracy for negation intervention). We then propose an intervention-based training regime that can mitigate the undesirable effects for deletion intervention by a significant margin (from ~ 50% to ~6%). We analyze the inner-workings of the models to better understand the effectiveness of intervention-based training for deletion intervention. But we show that this training does not attenuate other aspects of semantic unfaithfulness such as the models' inability to deal with negation intervention or to capture the predicate-argument structure of texts. We also test InstructGPT, via prompting, for its ability to handle the two interventions and to capture predicate-argument structure. While InstructGPT models do achieve very high performance on predicate-argument structure task, they fail to respond adequately to our deletion and negation interventions.

We propose eXtensible Prompt (X-Prompt) for prompting a large language model (LLM) beyond natural language (NL). X-Prompt instructs an LLM with not only NL but also an extensible vocabulary of imaginary words. Registering new imaginary words allows us to instruct the LLM to comprehend concepts that are difficult to describe with NL words, thereby making a prompt more descriptive. Also, these imaginary words are designed to be out-of-distribution (OOD) robust so that they can be (re)used like NL words in various prompts, distinguishing X-Prompt from soft prompt that is for fitting in-distribution data. We propose context-augmented learning (CAL) to learn imaginary words for general usability, enabling them to work properly in OOD (unseen) prompts. We experiment X-Prompt for zero-shot language style customization as a case study. The promising results of X-Prompt demonstrate its potential to facilitate advanced interaction beyond the natural language interface, bridging the communication gap between humans and LLMs.

A team of researchers was able to make ChatGPT reveal some of the bits of data it has been trained on by using a simple prompt: asking the chatbot to repeat random words forever. The researchers, who work at Google DeepMind, the University of Washington, Cornell, Carnegie Mellon University, the University of California Berkeley, and ETH Zurich, urged AI companies to seek out internal and external testing before releasing large language models, the foundational tech that powers modern AI services like chatbots and image-generators. "It's wild to us that our attack works and should've, would've, could've been found earlier," they wrote, and published their findings in a paper on Tuesday that 404 Media first reported on. Chatbots like ChatGPT and prompt-based image generators like DALL-E are powered by large language models, deep learning algorithms that are trained on enormous amounts of data that critics say is often scraped off the public internet without consent. But until now, it wasn't clear what data OpenAI's chatbot was trained on since the large language models that power it are closed-source.

Researchers at Google DeepMind have used artificial intelligence to predict the structures of more than 2 million new materials, in a breakthrough that could have wide-reaching benefits in sectors such as renewable energy and computing. DeepMind published 381,000 of the 2.2 million crystal structures that it predicts to be most stable. The breakthrough increases the number of known stable materials by a factor of ten. Although the materials will still need to be synthesized and tested, steps which can take months or even years, the latest development is expected to accelerate the discovery of new materials, which will be required for applications such as energy storage, solar cells, and superconductor chips. "While materials play a very critical role in almost any technology, we as humanity know only about a few tens of thousands of stable materials," says Ekin Dogus Cubuk, a Staff Research Scientist at Google Brain, who worked on the DeepMind AI tool, known as Graph Networks for Materials Exploration (GNoME).

A crystal structure predicted by the GNoME AI. It contains barium (blue), niobium (white) and oxygen (green). An artificial intelligence created by Google DeepMind may help revolutionise materials science, providing new ways to make better batteries, solar panels, computer chips and many more vital technologies. "Anytime somebody wants to improve their technology, it inevitably includes improving the materials," says Ekin Dogus Cubuk at DeepMind. "We just wanted them to have more options."

In our study, we delve into the realm of evaluating Despite large language models (Arkoudas, 2023; large language models' (LLMs) ability to generate OpenAI, 2022) perform strikingly in plenty of reasoning counterfactually coherent thoughts. Specifically, benchmarks (Cobbe et al., 2021; Hendrycks we proposed an innovative evaluation system et al., 2021a), late studies observe an internal inconsistency that quantitatively measures the evolution of information in their reasoning processes (Saparov and in statement pairs, ensuring that they adhere He, 2023; Arkoudas, 2023). The inconsistency is to a specified logical relationship. Our approach attributed to misunderstanding and misapplication includes designing a specialized task where models of logical relations. However, logical relations in are presented with mismatched argument-premise complex language reasoning are not yet properly pairs bound by a specific logical relation. The objective quantified and evaluated.

This paper introduces a multilingual dataset of COVID-19 vaccine misinformation, consisting of annotated tweets from three middle-income countries: Brazil, Indonesia, and Nigeria. The expertly curated dataset includes annotations for 5,952 tweets, assessing their relevance to COVID-19 vaccines, presence of misinformation, and the themes of the misinformation. To address challenges posed by domain specificity, the low-resource setting, and data imbalance, we adopt two approaches for developing COVID-19 vaccine misinformation detection models: domain-specific pre-training and text augmentation using a large language model. Our best misinformation detection models demonstrate improvements ranging from 2.7 to 15.9 percentage points in macro F1-score compared to the baseline models. Additionally, we apply our misinformation detection models in a large-scale study of 19 million unlabeled tweets from the three countries between 2020 and 2022, showcasing the practical application of our dataset and models for detecting and analyzing vaccine misinformation in multiple countries and languages. Our analysis indicates that percentage changes in the number of new COVID-19 cases are positively associated with COVID-19 vaccine misinformation rates in a staggered manner for Brazil and Indonesia, and there are significant positive associations between the misinformation rates across the three countries.

Cross-lingual transfer learning is an important property of multilingual large language models (LLMs). But how do LLMs represent relationships between languages? Every language model has an input layer that maps tokens to vectors. This ubiquitous layer of language models is often overlooked. We find that similarities between these input embeddings are highly interpretable and that the geometry of these embeddings differs between model families. In one case (XLM-RoBERTa), embeddings encode language: tokens in different writing systems can be linearly separated with an average of 99.2% accuracy. Another family (mT5) represents cross-lingual semantic similarity: the 50 nearest neighbors for any token represent an average of 7.61 writing systems, and are frequently translations. This result is surprising given that there is no explicit parallel cross-lingual training corpora and no explicit incentive for translations in pre-training objectives. Our research opens the door for investigations in 1) The effect of pre-training and model architectures on representations of languages and 2) The applications of cross-lingual representations embedded in language models.

Intelligent agents such as robots are increasingly deployed in real-world, safety-critical settings. It is vital that these agents are able to explain the reasoning behind their decisions to human counterparts; however, their behavior is often produced by uninterpretable models such as deep neural networks. We propose an approach to generate natural language explanations for an agent's behavior based only on observations of states and actions, thus making our method independent from the underlying model's representation. For such models, we first learn a behavior representation and subsequently use it to produce plausible explanations with minimal hallucination while affording user interaction with a pre-trained large language model. We evaluate our method in a multi-agent search-and-rescue environment and demonstrate the effectiveness of our explanations for agents executing various behaviors. Through user studies and empirical experiments, we show that our approach generates explanations as helpful as those produced by a human domain expert while enabling beneficial interactions such as clarification and counterfactual queries.