Following on recent advances in large language models (LLMs) and subsequent chat models, a new wave of large vision-language models (LVLMs) has emerged. Such models can incorporate images as input in addition to text, and perform tasks such as visual question answering, image captioning, story generation, etc. Here, we examine potential gender and racial biases in such systems, based on the perceived characteristics of the people in the input images. To accomplish this, we present a new dataset PAIRS (PArallel Images for eveRyday Scenarios). The PAIRS dataset contains sets of AI-generated images of people, such that the images are highly similar in terms of background and visual content, but differ along the dimensions of gender (man, woman) and race (Black, white). By querying the LVLMs with such images, we observe significant differences in the responses according to the perceived gender or race of the person depicted.

Aligning large language models (LLMs) with human values is imperative to mitigate potential adverse effects resulting from their misuse. Drawing from the sociological insight that acknowledging all parties' concerns is a key factor in shaping human values, this paper proposes a novel direction to align LLMs by themselves: social scene simulation. To achieve this, we present MATRIX, a novel social scene simulator that emulates realistic scenes around a user's input query, enabling the LLM to take social consequences into account before responding. MATRIX serves as a virtual rehearsal space, akin to a Monopolylogue, where the LLM performs diverse roles related to the query and practice by itself. To inject this alignment, we fine-tune the LLM with MATRIX-simulated data, ensuring adherence to human values without compromising inference speed. We theoretically show that the LLM with MATRIX outperforms Constitutional AI under mild assumptions. Finally, extensive experiments validate that our method outperforms over 10 baselines across 4 benchmarks. As evidenced by 875 user ratings, our tuned 13B-size LLM exceeds GPT-4 in aligning with human values.

This work investigates adversarial training in the context of margin-based linear classifiers in the high-dimensional regime where the dimension $d$ and the number of data points $n$ diverge with a fixed ratio $\alpha = n / d$. We introduce a tractable mathematical model where the interplay between the data and adversarial attacker geometries can be studied, while capturing the core phenomenology observed in the adversarial robustness literature. Our main theoretical contribution is an exact asymptotic description of the sufficient statistics for the adversarial empirical risk minimiser, under generic convex and non-increasing losses. Our result allow us to precisely characterise which directions in the data are associated with a higher generalisation/robustness trade-off, as defined by a robustness and a usefulness metric. In particular, we unveil the existence of directions which can be defended without penalising accuracy. Finally, we show the advantage of defending non-robust features during training, identifying a uniform protection as an inherently effective defence mechanism.

Misuse of the Large Language Models (LLMs) has raised widespread concern. To address this issue, safeguards have been taken to ensure that LLMs align with social ethics. However, recent findings have revealed an unsettling vulnerability bypassing the safeguards of LLMs, known as jailbreak attacks. By applying techniques, such as employing role-playing scenarios, adversarial examples, or subtle subversion of safety objectives as a prompt, LLMs can produce an inappropriate or even harmful response. While researchers have studied several categories of jailbreak attacks, they have done so in isolation. To fill this gap, we present the first large-scale measurement of various jailbreak attack methods. We concentrate on 13 cutting-edge jailbreak methods from four categories, 160 questions from 16 violation categories, and six popular LLMs. Our extensive experimental results demonstrate that the optimized jailbreak prompts consistently achieve the highest attack success rates, as well as exhibit robustness across different LLMs. Some jailbreak prompt datasets, available from the Internet, can also achieve high attack success rates on many LLMs, such as ChatGLM3, GPT-3.5, and PaLM2. Despite the claims from many organizations regarding the coverage of violation categories in their policies, the attack success rates from these categories remain high, indicating the challenges of effectively aligning LLM policies and the ability to counter jailbreak attacks. We also discuss the trade-off between the attack performance and efficiency, as well as show that the transferability of the jailbreak prompts is still viable, becoming an option for black-box models. Overall, our research highlights the necessity of evaluating different jailbreak methods. We hope our study can provide insights for future research on jailbreak attacks and serve as a benchmark tool for evaluating them for practitioners.

The encounter situation between marine vessels determines how they should navigate to obey COLREGs, but time-varying and stochastic uncertainty in estimation of angles of encounter, and of closest point of approach, easily give rise to different assessment of situation at two approaching vessels. This may lead to high-risk conditions and could cause collision. This article considers decision making under uncertainty and suggests a novel method for probabilistic interpretation of vessel encounters that is explainable and provides a measure of uncertainty in the evaluation. The method is equally useful for decision support on a manned bridge as on Marine Autonomous Surface Ships (MASS) where it provides input for automated navigation. The method makes formal safety assessment and validation feasible. We obtain a resilient algorithm for machine interpretation of COLREGs under uncertainty and show its efficacy by simulations.

Long-form generations from large language models (LLMs) contains a mix of factual and non-factual claims, making evaluating factuality difficult. To evaluate factual precision of long-form generations in a more fine-grained way, prior works propose to decompose long-form generations into multiple verifiable facts and verify those facts independently. The factuality of the generation is the proportion of verifiable facts among all the facts. Such methods assume that combining factual claims forms a factual paragraph. This paper shows that the assumption can be violated due to entity ambiguity. We show that LLMs can generate paragraphs that contain verifiable facts, but the facts are combined to form a non-factual paragraph due to entity ambiguity. We further reveal that existing factual precision metrics, including FActScore and citation recall, cannot properly evaluate the factuality of these non-factual paragraphs. To address this, we introduce an enhanced metric, D-FActScore, specifically designed for content with ambiguous entities. We evaluate the D-FActScores of people biographies generated with retrieval-augmented generation (RAG). We show that D-FActScore can better assess the factuality of paragraphs with entity ambiguity than FActScore. We also find that four widely used open-source LLMs tend to mix information of distinct entities to form non-factual paragraphs.

Recent advancements in federated learning (FL) have produced models that retain user privacy by training across multiple decentralized devices or systems holding local data samples. However, these strategies often neglect the inherent challenges of statistical heterogeneity and vulnerability to adversarial attacks, which can degrade model robustness and fairness. Personalized FL strategies offer some respite by adjusting models to fit individual client profiles, yet they tend to neglect server-side aggregation vulnerabilities. To address these issues, we propose Reinforcement Federated Learning (RFL), a novel framework that leverages deep reinforcement learning to adaptively optimize client contribution during aggregation, thereby enhancing both model robustness against malicious clients and fairness across participants under non-identically distributed settings. To achieve this goal, we propose a meticulous approach involving a Deep Deterministic Policy Gradient-based algorithm for continuous control of aggregation weights, an innovative client selection method based on model parameter distances, and a reward mechanism guided by validation set performance. Empirically, extensive experiments demonstrate that, in terms of robustness, RFL outperforms the state-of-the-art methods, while maintaining comparable levels of fairness, offering a promising solution to build resilient and fair federated systems.

In this work, we investigate to use Large Language Models (LLMs) for rumor detection on social media. However, it is challenging for LLMs to reason over the entire propagation information on social media, which contains news contents and numerous comments, due to LLMs may not concentrate on key clues in the complex propagation information, and have trouble in reasoning when facing massive and redundant information. Accordingly, we propose an LLM-empowered Rumor Detection (LeRuD) approach, in which we design prompts to teach LLMs to reason over important clues in news and comments, and divide the entire propagation information into a Chain-of-Propagation for reducing LLMs' burden. We conduct extensive experiments on the Twitter and Weibo datasets, and LeRuD outperforms several state-of-the-art rumor detection models by 3.2% to 7.7%. Meanwhile, by applying LLMs, LeRuD requires no data for training, and thus shows more promising rumor detection ability in few-shot or zero-shot scenarios.

Moreover, we reproduce 10 representative In recent years, adversarial attacks have emerged as a methods of transferable adversarial attacks and integrated significant research direction for artificial intelligence and these methods into an open-source benchmark framework, machine learning, especially in the context of the security which is published on GitHub TAA-Bench, facilitating related of deep learning. It originates from the observation that deep researches. Our main contributions are: neural networks (DNNs) are sensitive to subtle perturbations in input data. Even imperceptible to the human eye, such changes We thoroughly collate existing methods of transferable can lead to incorrect output results [3]. Adversarial attacks adversarial attacks, and systematically analyse their can be categorized into two types based on the availability implementation principles. of model data: white-box attacks and black-box attacks [1], [18], [6], [19]. White-box attacks assume the model's internal We present an extensible, modular and open-source information are accessible, such as its parameters, structure, benchmark TAA-Bench that includes implementations and training data. In contrast, black-box attacks occur without of different types of transferable adversarial attacks to knowledge of the internal information of the attacked model, facilitate research and development in this field.

This paper will examine what makes a being intelligent, whether that be a biological being or an artificial silicon being on a computer. Special attention will be paid to the being having the ability to characterize and control a collective system of many identical conservative sub-systems conservatively interacting. The essence of intelligence will be found to be the golden rule -- "the collective acts as one" or "knowing the global consequences of local actions". The flow of the collective is a small set of twinkling textures, that are governed by a puppeteer who is pulling a small number of strings according to a geodesic motion of least action, determined by the symmetries. Controlling collective conservative systems is difficult and has historically been done by adding significant viscosity to the system to stabilize the desirable meta stable equilibriums of maximum performance, but it degrades or destroys them in the process. There is an alternative. Once the optimum twinkling textures of the meta stable equilibriums are identified, the collective system can be moved to the optimum twinkling textures, then quickly vibrated according to the textures so that the collective system remains at the meta stable equilibrium. Well educated intelligence knows the global consequences of its local actions so that it will not take short term actions that will lead to poor long term outcomes. In contrast, trained intelligence or trained stupidity will optimize its short term actions, leading to poor long term outcomes. Well educated intelligence is inherently good, but trained stupidity is inherently evil and should be feared. Particular attention is paid to the control and optimization of economic and social collectives. These new results are also applicable to physical collectives such as fields, fluids and plasmas.