Generative artificial intelligence (AI) technologies and large models are producing realistic outputs across various domains, such as images, text, speech, and music. Creating these advanced generative models requires significant resources, particularly large and high-quality datasets. To minimize training expenses, many algorithm developers use data created by the models themselves as a cost-effective training solution. However, not all synthetic data effectively improve model performance, necessitating a strategic balance in the use of real versus synthetic data to optimize outcomes. Currently, the previously well-controlled integration of real and synthetic data is becoming uncontrollable. The widespread and unregulated dissemination of synthetic data online leads to the contamination of datasets traditionally compiled through web scraping, now mixed with unlabeled synthetic data. This trend portends a future where generative AI systems may increasingly rely blindly on consuming self-generated data, raising concerns about model performance and ethical issues. What will happen if generative AI continuously consumes itself without discernment? What measures can we take to mitigate the potential adverse effects? There is a significant gap in the scientific literature regarding the impact of synthetic data use in generative AI, particularly in terms of the fusion of multimodal information. To address this research gap, this review investigates the consequences of integrating synthetic data blindly on training generative AI on both image and text modalities and explores strategies to mitigate these effects. The goal is to offer a comprehensive view of synthetic data's role, advocating for a balanced approach to its use and exploring practices that promote the sustainable development of generative AI technologies in the era of large models.

Information pooling has been extensively formalised across various logical frameworks in distributed systems [1,30,13,8], characterized by diverse information-sharing patterns. These approaches generally adopt an intersection perspective, aggregating all possible information, regardless of whether it is known or unknown to the agents. In contrast, this work adopts a unique stance, emphasising that sharing knowledge means distributing what is known, rather than what remains uncertain. This paper introduces new modal logics for knowledge pooling and sharing, ranging from a novel language of knowledge pooling to a dynamic mechanism for knowledge sharing. It also outlines their axiomatizations and discusses a potential framework for permissible knowledge pooling. Keywords: Information Pooling, Distributed Knowledge, Knowledge Pooling, Permissible Knowledge Sharing. What disclosures are considered sensitive information leaks? What messages can be shared with a cooperative partner without breaching confidentiality agreements?

The Online Safety Act in the UK is, quietly, one of the most important pieces of legislation to have come out of this government. Admittedly, the competition is slim. But as time goes by, and more and more of the act begins to take effect, we're starting to see how it will reshape the internet. Social media firms have been told to "tame aggressive algorithms" that recommend harmful content to children, as part of Ofcom's new safety codes of practice. The children's safety codes, introduced as part of the Online Safety Act, let Ofcom set new, tight rules for internet companies and how they can interact with children.

Large language models (LLMs) have shown success in many natural language processing tasks. Despite rigorous safety alignment processes, supposedly safety-aligned LLMs like Llama 2 and Claude 2 are still susceptible to jailbreaks, leading to security risks and abuse of the models. One option to mitigate such risks is to augment the LLM with a dedicated "safeguard", which checks the LLM's inputs or outputs for undesired behaviour. A promising approach is to use the LLM itself as the safeguard. Nonetheless, baseline methods, such as prompting the LLM to self-classify toxic content, demonstrate limited efficacy. We hypothesise that this is due to domain shift: the alignment training imparts a self-censoring behaviour to the model ("Sorry I can't do that"), while the self-classify approach shifts it to a classification format ("Is this prompt malicious"). In this work, we propose PARDEN, which avoids this domain shift by simply asking the model to repeat its own outputs. PARDEN neither requires finetuning nor white box access to the model. We empirically verify the effectiveness of our method and show that PARDEN significantly outperforms existing jailbreak detection baselines for Llama-2 and Claude-2. Code and data are available at https://github.com/Ed-Zh/PARDEN. We find that PARDEN is particularly powerful in the relevant regime of high True Positive Rate (TPR) and low False Positive Rate (FPR). For instance, for Llama2-7B, at TPR equal to 90%, PARDEN accomplishes a roughly 11x reduction in the FPR from 24.8% to 2.0% on the harmful behaviours dataset.

Machine learning (ML)-based content moderation tools are essential to keep online spaces free from hateful communication. Yet, ML tools can only be as capable as the quality of the data they are trained on allows them. While there is increasing evidence that they underperform in detecting hateful communications directed towards specific identities and may discriminate against them, we know surprisingly little about the provenance of such bias. To fill this gap, we present a systematic review of the datasets for the automated detection of hateful communication introduced over the past decade, and unpack the quality of the datasets in terms of the identities that they embody: those of the targets of hateful communication that the data curators focused on, as well as those unintentionally included in the datasets. We find, overall, a skewed representation of selected target identities and mismatches between the targets that research conceptualizes and ultimately includes in datasets. Yet, by contextualizing these findings in the language and location of origin of the datasets, we highlight a positive trend towards the broadening and diversification of this research space.

This paper explores the potential of large language models (LLMs) to make the Aeronautical Regulations of Colombia (RAC) more accessible. Given the complexity and extensive technicality of the RAC, this study introduces a novel approach to simplifying these regulations for broader understanding. By developing the first-ever RAC database, which contains 24,478 expertly labeled question-and-answer pairs, and fine-tuning LLMs specifically for RAC applications, the paper outlines the methodology for dataset assembly, expert-led annotation, and model training. Utilizing the Gemma1.1 2b model along with advanced techniques like Unsloth for efficient VRAM usage and flash attention mechanisms, the research aims to expedite training processes. This initiative establishes a foundation to enhance the comprehensibility and accessibility of RAC, potentially benefiting novices and reducing dependence on expert consultations for navigating the aviation industry's regulatory landscape. You can visit the dataset (https://huggingface.co/somosnlp/gemma-1.1-2b-it_ColombiaRAC_FullyCurated_format_chatML_V1) and the model (https://huggingface.co/datasets/somosnlp/ColombiaRAC_FullyCurated) here.

The rapid advancement of large language models (LLMs) has made it increasingly difficult to distinguish between text written by humans and machines. Addressing this, we propose a novel method for generating watermarks that strategically alters token probabilities during generation. Unlike previous works, this method uniquely employs linguistic features such as stylometry. Concretely, we introduce acrostica and sensorimotor norms to LLMs. Further, these features are parameterized by a key, which is updated every sentence. To compute this key, we use semantic zero shot classification, which enhances resilience. In our evaluation, we find that for three or more sentences, our method achieves a false positive and false negative rate of 0.02. For the case of a cyclic translation attack, we observe similar results for seven or more sentences. This research is of particular of interest for proprietary LLMs to facilitate accountability and prevent societal harm.

The SemEval task on Argument Reasoning in Civil Procedure is challenging in that it requires understanding legal concepts and inferring complex arguments. Currently, most Large Language Models (LLM) excelling in the legal realm are principally purposed for classification tasks, hence their reasoning rationale is subject to contention. The approach we advocate involves using a powerful teacher-LLM (ChatGPT) to extend the training dataset with explanations and generate synthetic data. The resulting data are then leveraged to fine-tune a small student-LLM. Contrary to previous work, our explanations are not directly derived from the teacher's internal knowledge. Instead they are grounded in authentic human analyses, therefore delivering a superior reasoning signal. Additionally, a new `mutation' method generates artificial data instances inspired from existing ones. We are publicly releasing the explanations as an extension to the original dataset, along with the synthetic dataset and the prompts that were used to generate both. Our system ranked 15th in the SemEval competition. It outperforms its own teacher and can produce explanations aligned with the original human analyses, as verified by legal experts.

The current safeguard mechanisms for large language models (LLMs) are indeed susceptible to jailbreak attacks, making them inherently fragile. Even the process of fine-tuning on apparently benign data for downstream tasks can jeopardize safety. One potential solution is to conduct safety fine-tuning subsequent to downstream fine-tuning. However, there's a risk of catastrophic forgetting during safety fine-tuning, where LLMs may regain safety measures but lose the task-specific knowledge acquired during downstream fine-tuning. In this paper, we introduce a safety realignment framework through subspace-oriented model fusion (SOMF), aiming to combine the safeguard capabilities of initially aligned model and the current fine-tuned model into a realigned model. Our approach begins by disentangling all task vectors from the weights of each fine-tuned model. We then identify safety-related regions within these vectors by subspace masking techniques. Finally, we explore the fusion of the initial safely aligned LLM with all task vectors based on the identified safety subspace. We validate that our safety realignment framework satisfies the safety requirements of a single fine-tuned model as well as multiple models during their fusion. Our findings confirm that SOMF preserves safety without notably compromising performance on downstream tasks, including instruction following in Chinese, English, and Hindi, as well as problem-solving capabilities in Code and Math.

What can contemporary machine learning (ML) models do? Given the proliferation of ML models in society, answering this question matters to a variety of stakeholders, both public and private. The evaluation of models' capabilities is rapidly emerging as a key subfield of modern ML, buoyed by regulatory attention and government grants. Despite this, the notion of an ML model possessing a capability has not been interrogated: what are we saying when we say that a model is able to do something? And what sorts of evidence bear upon this question? In this paper, we aim to answer these questions, using the capabilities of large language models (LLMs) as a running example. Drawing on the large philosophical literature on abilities, we develop an account of ML models' capabilities which can be usefully applied to the nascent science of model evaluation. Our core proposal is a conditional analysis of model abilities (CAMA): crudely, a machine learning model has a capability to X just when it would reliably succeed at doing X if it 'tried'. The main contribution of the paper is making this proposal precise in the context of ML, resulting in an operationalisation of CAMA applicable to LLMs. We then put CAMA to work, showing that it can help make sense of various features of ML model evaluation practice, as well as suggest procedures for performing fair inter-model comparisons.