The AI act is the European Union-wide regulation of AI systems. It includes specific provisions for general-purpose AI models which however need to be further interpreted in terms of technical standards and state-of-art studies to ensure practical compliance solutions. This paper examines the AI act requirements for providers and deployers of general-purpose AI and further proposes uncertainty estimation as a suitable measure for legal compliance and quality assurance in training of such models. We argue that uncertainty estimation should be a required component for deploying models in the real world, and under the EU AI Act, it could fulfill several requirements for transparency, accuracy, and trustworthiness. However, generally using uncertainty estimation methods increases the amount of computation, producing a dilemma, as computation might go over the threshold ($10^{25}$ FLOPS) to classify the model as a systemic risk system which bears more regulatory burden.

We introduce a group of related methods for binary classification tasks using probes of the hidden state activations in large language models (LLMs). Performance is on par with the largest and most advanced LLMs currently available, but requiring orders of magnitude fewer computational resources and not requiring labeled data. This approach involves translating class labels into a semantically rich description, spontaneous symmetry breaking of multilayer perceptron probes for unsupervised learning and inference, training probes to generate confidence scores (prior probabilities) from hidden state activations subject to known constraints via entropy maximization, and selecting the most confident probe model from an ensemble for prediction. These techniques are evaluated on four datasets using five base LLMs.

Detecting out-of-distribution (OOD) data is crucial in machine learning applications to mitigate the risk of model overconfidence, thereby enhancing the reliability and safety of deployed systems. The majority of existing OOD detection methods predominantly address uni-modal inputs, such as images or texts. In the context of multi-modal documents, there is a notable lack of extensive research on the performance of these methods, which have primarily been developed with a focus on computer vision tasks. We propose a novel methodology termed as attention head masking (AHM) for multi-modal OOD tasks in document classification systems. Our empirical results demonstrate that the proposed AHM method outperforms all state-of-the-art approaches and significantly decreases the false positive rate (FPR) compared to existing solutions up to 7.5\%. This methodology generalizes well to multi-modal data, such as documents, where visual and textual information are modeled under the same Transformer architecture. To address the scarcity of high-quality publicly available document datasets and encourage further research on OOD detection for documents, we introduce FinanceDocs, a new document AI dataset. Our code and dataset are publicly available.

Competitive debate is a complex task of computational argumentation. Large Language Models (LLMs) suffer from hallucinations and lack competitiveness in this field. To address these challenges, we introduce Agent for Debate (Agent4Debate), a dynamic multi-agent framework based on LLMs designed to enhance their capabilities in competitive debate. Drawing inspiration from human behavior in debate preparation and execution, Agent4Debate employs a collaborative architecture where four specialized agents, involving Searcher, Analyzer, Writer, and Reviewer, dynamically interact and cooperate. These agents work throughout the debate process, covering multiple stages from initial research and argument formulation to rebuttal and summary. To comprehensively evaluate framework performance, we construct the Competitive Debate Arena, comprising 66 carefully selected Chinese debate motions. We recruit ten experienced human debaters and collect records of 200 debates involving Agent4Debate, baseline models, and humans. The evaluation employs the Debatrix automatic scoring system and professional human reviewers based on the established Debatrix-Elo and Human-Elo ranking. Experimental results indicate that the state-of-the-art Agent4Debate exhibits capabilities comparable to those of humans. Furthermore, ablation studies demonstrate the effectiveness of each component in the agent structure.

In today's era, where large language models (LLMs) are integrated into numerous real-world applications, ensuring their safety and robustness is crucial for responsible AI usage. Automated red-teaming methods play a key role in this process by generating adversarial attacks to identify and mitigate potential vulnerabilities in these models. However, existing methods often struggle with slow performance, limited categorical diversity, and high resource demands. While Rainbow Teaming, a recent approach, addresses the diversity challenge by framing adversarial prompt generation as a quality-diversity search, it remains slow and requires a large fine-tuned mutator for optimal performance. To overcome these limitations, we propose Ferret, a novel approach that builds upon Rainbow Teaming by generating multiple adversarial prompt mutations per iteration and using a scoring function to rank and select the most effective adversarial prompt. We explore various scoring functions, including reward models, Llama Guard, and LLM-as-a-judge, to rank adversarial mutations based on their potential harm to improve the efficiency of the search for harmful mutations. Our results demonstrate that Ferret, utilizing a reward model as a scoring function, improves the overall attack success rate (ASR) to 95%, which is 46% higher than Rainbow Teaming. Additionally, Ferret reduces the time needed to achieve a 90% ASR by 15.2% compared to the baseline and generates adversarial prompts that are transferable i.e. effective on other LLMs of larger size. Our codes are available at https://github.com/declare-lab/ferret.

With the emergence of large language models (LLMs), LLM-powered multi-agent systems (LLM-MA systems) have been proposed to tackle real-world tasks. However, their agents mostly follow predefined Standard Operating Procedures (SOPs) that remain unchanged across the whole interaction, lacking autonomy and scalability. Additionally, current solutions often overlook the necessity for effective agent cooperation. To address the above limitations, we propose MegaAgent, a practical framework designed for autonomous cooperation in large-scale LLM Agent systems. MegaAgent leverages the autonomy of agents to dynamically generate agents based on task requirements, incorporating features such as automatically dividing tasks, systematic planning and monitoring of agent activities, and managing concurrent operations. In addition, MegaAgent is designed with a hierarchical structure and employs system-level parallelism to enhance performance and boost communication. We demonstrate the effectiveness of MegaAgent through Gobang game development, showing that it outperforms popular LLM-MA systems; and national policy simulation, demonstrating its high autonomy and potential to rapidly scale up to 590 agents while ensuring effective cooperation among them. Our results indicate that MegaAgent is the first autonomous large-scale LLM-MA system with no pre-defined SOPs, high effectiveness and scalability, paving the way for further research in this field. Our code is at https://anonymous.4open.science/r/MegaAgent-81F3.

Due to their length and complexity, long regulatory texts are challenging to summarize. To address this, a multi-step extractive-abstractive architecture is proposed to handle lengthy regulatory documents more effectively. In this paper, we show that the effectiveness of a two-step architecture for summarizing long regulatory texts varies significantly depending on the model used. Specifically, the two-step architecture improves the performance of decoder-only models. For abstractive encoder-decoder models with short context lengths, the effectiveness of an extractive step varies, whereas for long-context encoder-decoder models, the extractive step worsens their performance. This research also highlights the challenges of evaluating generated texts, as evidenced by the differing results from human and automated evaluations. Most notably, human evaluations favoured language models pretrained on legal text, while automated metrics rank general-purpose language models higher. The results underscore the importance of selecting the appropriate summarization strategy based on model architecture and context length.

This study investigates the effects of occlusions on the fairness of face recognition systems, particularly focusing on demographic biases. Using the Racial Faces in the Wild (RFW) dataset and synthetically added realistic occlusions, we evaluate their effect on the performance of face recognition models trained on the BUPT-Balanced and BUPT-GlobalFace datasets. We note increases in the dispersion of FMR, FNMR, and accuracy alongside decreases in fairness according to Equilized Odds, Demographic Parity, STD of Accuracy, and Fairness Discrepancy Rate. Additionally, we utilize a pixel attribution method to understand the importance of occlusions in model predictions, proposing a new metric, Face Occlusion Impact Ratio (FOIR), that quantifies the extent to which occlusions affect model performance across different demographic groups. Our results indicate that occlusions exacerbate existing demographic biases, with models placing higher importance on occlusions in an unequal fashion, particularly affecting African individuals more severely.

There has been a growing interest in multi-agent simulators in the domain of economic modeling. However, contemporary research often involves developing reinforcement learning (RL) based models that focus solely on a single type of agents, such as households, firms, or the government. Such an approach overlooks the adaptation of interacting agents thereby failing to capture the complexity of real-world economic systems. In this work, we consider a multi-agent simulator comprised of RL agents of numerous types, including heterogeneous households, firm, central bank and government. In particular, we focus on the crucial role of the government in distributing tax credits to households. We conduct two broad categories of comprehensive experiments dealing with the impact of tax credits on 1) households with varied degrees of myopia (short-sightedness in spending and saving decisions), and 2) households with diverse liquidity profiles. The first category of experiments examines the impact of the frequency of tax credits (e.g. annual vs quarterly) on consumption patterns of myopic households. The second category of experiments focuses on the impact of varying tax credit distribution strategies on households with differing liquidities. We validate our simulation model by reproducing trends observed in real households upon receipt of unforeseen, uniform tax credits, as documented in a JPMorgan Chase report. Based on the results of the latter, we propose an innovative tax credit distribution strategy for the government to reduce inequality among households. We demonstrate the efficacy of this strategy in improving social welfare in our simulation results.

Detecting and quantifying issue framing in textual discourse - the perspective one takes to a given topic (e.g. climate science vs. denialism, misogyny vs. gender equality) - is highly valuable to a range of end-users from social and political scientists to program evaluators and policy analysts. However, conceptual framing is notoriously challenging for automated natural language processing (NLP) methods since the words and phrases used by either `side' of an issue are often held in common, with only subtle stylistic flourishes separating their use. Here we develop and rigorously evaluate new detection methods for issue framing and narrative analysis within large text datasets. By introducing a novel application of next-token log probabilities derived from generative large language models (LLMs) we show that issue framing can be reliably and efficiently detected in large corpora with only a few examples of either perspective on a given issue, a method we call `paired completion'. Through 192 independent experiments over three novel, synthetic datasets, we evaluate paired completion against prompt-based LLM methods and labelled methods using traditional NLP and recent LLM contextual embeddings. We additionally conduct a cost-based analysis to mark out the feasible set of performant methods at production-level scales, and a model bias analysis. Together, our work demonstrates a feasible path to scalable, accurate and low-bias issue-framing in large corpora.