The rapid development of advanced AI agents and the imminent deployment of many instances of these agents will give rise to multi-agent systems of unprecedented complexity. These systems pose novel and under-explored risks. In this report, we provide a structured taxonomy of these risks by identifying three key failure modes (miscoordination, conflict, and collusion) based on agents' incentives, as well as seven key risk factors (information asymmetries, network effects, selection pressures, destabilising dynamics, commitment problems, emergent agency, and multi-agent security) that can underpin them. We highlight several important instances of each risk, as well as promising directions to help mitigate them. By anchoring our analysis in a range of real-world examples and experimental evidence, we illustrate the distinct challenges posed by multi-agent systems and their implications for the safety, governance, and ethics of advanced AI.

Modern text classification methods heavily rely on contextual embeddings from large language models (LLMs). Compared to human-engineered features, these embeddings provide automatic and effective representations for classification model training. However, they also introduce a challenge: we lose the ability to manually remove unintended features, such as sensitive or task-irrelevant features, to guarantee regulatory compliance or improve the generalizability of classification models. This limitation arises because LLM embeddings are opaque and difficult to interpret. In this paper, we propose a novel framework to identify and regularize unintended features in the LLM latent space. Specifically, we first pre-train a sparse autoencoder (SAE) to extract interpretable features from LLM latent spaces. To ensure the SAE can capture task-specific features, we further fine-tune it on task-specific datasets. In training the classification model, we propose a simple and effective regularizer, by minimizing the similarity between the classifier weights and the identified unintended feature, to remove the impacts of these unintended features toward classification. We evaluate the proposed framework on three real-world tasks, including toxic chat detection, reward modeling, and disease diagnosis. Results show that the proposed framework can significantly improve the classifier's generalizability by regularizing those features that are not semantically correlated to each task. This work pioneers controllable text classification on LLM latent spaces by leveraging interpreted features to address generalizability, fairness, and privacy challenges. We will release our code and data once accepted.

Two commonly-employed strategies to combat the rise of misinformation on social media are (i) fact-checking by professional organisations and (ii) community moderation by platform users. Policy changes by Twitter/X and, more recently, Meta, signal a shift away from partnerships with fact-checking organisations and towards an increased reliance on crowdsourced community notes. However, the extent and nature of dependencies between fact-checking and helpful community notes remain unclear. To address these questions, we use language models to annotate a large corpus of Twitter/X community notes with attributes such as topic, cited sources, and whether they refute claims tied to broader misinformation narratives. Our analysis reveals that community notes cite fact-checking sources up to five times more than previously reported. Fact-checking is especially crucial for notes on posts linked to broader narratives, which are twice as likely to reference fact-checking sources compared to other sources. In conclusion, our results show that successful community moderation heavily relies on professional fact-checking.

Exploring new ideas is a fundamental aspect of research and development (R\&D), which often occurs in competitive environments. Most ideas are subsequent, i.e. one idea today leads to more ideas tomorrow. According to one approach, the best way to encourage exploration is by granting protection on discoveries to the first innovator. Correspondingly, only the one who made the first discovery can use the new knowledge and benefit from subsequent discoveries, which in turn should increase the initial motivation to explore. An alternative approach to promote exploration favors the \emph{sharing of knowledge} from discoveries among researchers allowing explorers to use each others' discoveries to develop further knowledge, as in the open-source community. With no protection, all explorers have access to all existing discoveries and new directions are explored faster. We present a game theoretic analysis of an abstract research-and-application game which clarifies the expected advantages and disadvantages of the two approaches under full information. We then compare the theoretical predictions with the observed behavior of actual players in the lab who operate under partial information conditions in both worlds. Our main experimental finding is that the no protection approach leads to \emph{more} investment efforts overall, in contrast to theoretical prediction and common economic wisdom, but in line with a familiar cognitive bias known as `underweighting of rare events'.

We devise an algorithm to generate sets of propositions that objectively instantiate graphs that support coherence-driven inference. We then benchmark the ability of large language models (LLMs) to reconstruct coherence graphs from (a straightforward transformation of) propositions expressed in natural language, with promising results from a single prompt to models optimized for reasoning. Combining coherence-driven inference with consistency evaluations by neural models may advance the state of the art in machine cognition.

We introduce Faina, the first dataset for fallacy detection that embraces multiple plausible answers and natural disagreement. Faina includes over 11K span-level annotations with overlaps across 20 fallacy types on social media posts in Italian about migration, climate change, and public health given by two expert annotators. Through an extensive annotation study that allowed discussion over multiple rounds, we minimize annotation errors whilst keeping signals of human label variation. Moreover, we devise a framework that goes beyond "single ground truth" evaluation and simultaneously accounts for multiple (equally reliable) test sets and the peculiarities of the task, i.e., partial span matches, overlaps, and the varying severity of labeling errors. Our experiments across four fallacy detection setups show that multi-task and multi-label transformer-based approaches are strong baselines across all settings. We release our data, code, and annotation guidelines to foster research on fallacy detection and human label variation more broadly.

Large Vision-Language Models (LVLMs) have recently gained attention due to their distinctive performance and broad applicability. While it has been previously shown that their efficacy in usage scenarios involving non-Western contexts falls short, existing studies are limited in scope, covering just a narrow range of cultures, focusing exclusively on a small number of cultural aspects, or evaluating a limited selection of models on a single task only. Towards globally inclusive LVLM research, we introduce GIMMICK, an extensive multimodal benchmark designed to assess a broad spectrum of cultural knowledge across 144 countries representing six global macro-regions. GIMMICK comprises six tasks built upon three new datasets that span 728 unique cultural events or facets on which we evaluated 20 LVLMs and 11 LLMs, including five proprietary and 26 open-weight models of all sizes. We systematically examine (1) regional cultural biases, (2) the influence of model size, (3) input modalities, and (4) external cues. Our analyses reveal strong biases toward Western cultures across models and tasks and highlight strong correlations between model size and performance, as well as the effectiveness of multimodal input and external geographic cues. We further find that models have more knowledge of tangible than intangible aspects (e.g., food vs. rituals) and that they excel in recognizing broad cultural origins but struggle with a more nuanced understanding.

With the increasing prevalence of recorded human speech, spoken language understanding (SLU) is essential for its efficient processing. In order to process the speech, it is commonly transcribed using automatic speech recognition technology. This speech-to-text transition introduces errors into the transcripts, which subsequently propagate to downstream NLP tasks, such as dialogue summarization. While it is known that transcript noise affects downstream tasks, a systematic approach to analyzing its effects across different noise severities and types has not been addressed. We propose a configurable framework for assessing task models in diverse noisy settings, and for examining the impact of transcript-cleaning techniques. The framework facilitates the investigation of task model behavior, which can in turn support the development of effective SLU solutions. We exemplify the utility of our framework on three SLU tasks and four task models, offering insights regarding the effect of transcript noise on tasks in general and models in particular. For instance, we find that task models can tolerate a certain level of noise, and are affected differently by the types of errors in the transcript.

Large language models (LLMs) are known to have the potential to generate harmful content, posing risks to users. While significant progress has been made in developing taxonomies for LLM risks and safety evaluation prompts, most studies have focused on monolingual contexts, primarily in English. However, language- and region-specific risks in bilingual contexts are often overlooked, and core findings can diverge from those in monolingual settings. In this paper, we introduce Qorgau, a novel dataset specifically designed for safety evaluation in Kazakh and Russian, reflecting the unique bilingual context in Kazakhstan, where both Kazakh (a low-resource language) and Russian (a high-resource language) are spoken. Experiments with both multilingual and language-specific LLMs reveal notable differences in safety performance, emphasizing the need for tailored, region-specific datasets to ensure the responsible and safe deployment of LLMs in countries like Kazakhstan. Warning: this paper contains example data that may be offensive, harmful, or biased.

This paper addresses the critical need for detecting bias in government documents, an underexplored area with significant implications for governance. Existing methodologies often overlook the unique context and far-reaching impacts of governmental documents, potentially obscuring embedded biases that shape public policy and citizen-government interactions. To bridge this gap, we introduce the Dutch Government Data for Bias Detection (DGDB), a dataset sourced from the Dutch House of Representatives and annotated for bias by experts. We fine-tune several BERT-based models on this dataset and compare their performance with that of generative language models. Additionally, we conduct a comprehensive error analysis that includes explanations of the models' predictions. Our findings demonstrate that fine-tuned models achieve strong performance and significantly outperform generative language models, indicating the effectiveness of DGDB for bias detection. This work underscores the importance of labeled datasets for bias detection in various languages and contributes to more equitable governance practices.