The off-switch problem is a critical challenge in AI control: if an AI system resists being switched off, it poses a significant risk. In this paper, we model the off-switch problem as a signalling game, where a human decision-maker communicates its preferences about some underlying decision problem to an AI agent, which then selects actions to maximise the human's utility. We assume that the human is a bounded rational agent and explore various bounded rationality mechanisms. Using real machine learning models, we reprove prior results and demonstrate that a necessary condition for an AI system to refrain from disabling its off-switch is its uncertainty about the human's utility. We also analyse how message costs influence optimal strategies and extend the analysis to scenarios involving incomparability.

Non-autoregressive (NAR) generative models are valuable because they can handle diverse conditional generation tasks in a more principled way than their autoregressive (AR) counterparts, which are constrained by sequential dependency requirements. Recent advancements in NAR models, such as diffusion language models, have demonstrated superior performance in unconditional generation compared to AR models (e.g., GPTs) of similar sizes. However, such improvements do not always lead to improved conditional generation performance. We show that a key reason for this gap is the difficulty in generalizing to conditional probability queries unseen during training. As a result, strong unconditional generation performance does not guarantee high-quality conditional generation. This paper proposes Tractable Transformers (Tracformer), a Transformer-based generative model that is more robust to different conditional generation tasks. Unlike existing models that rely solely on global contextual features derived from full inputs, Tracformers incorporate a sparse Transformer encoder to capture both local and global contextual information. This information is routed through a decoder for conditional generation. Empirical results demonstrate that Tracformers achieve state-of-the-art conditional generation performance on text modeling compared to recent diffusion and AR model baselines.

Achieving a delicate balance between fostering trust in law en- forcement and protecting the rights of both officers and civilians continues to emerge as a pressing research and product challenge in the world today. In the pursuit of fairness and transparency, this study presents an innovative AI-driven system designed to generate police report drafts from complex, noisy, and multi-role dialogue data. Our approach intelligently extracts key elements of law enforcement interactions and includes them in the draft, producing structured narratives that are not only high in quality but also reinforce accountability and procedural clarity. This frame- work holds the potential to transform the reporting process, ensur- ing greater oversight, consistency, and fairness in future policing practices. A demonstration video of our system can be accessed at https://drive.google.com/file/d/1kBrsGGR8e3B5xPSblrchRGj-Y-kpCHNO/view?usp=sharing

Large language models are increasingly used to process documents and facilitate question-answering on them. In our paper, we extract mechanistic circuits for this real-world language modeling task: context-augmented language modeling for extractive question-answering (QA) tasks and understand the potential benefits of circuits towards downstream applications such as data attribution to context information. We extract circuits as a function of internal model components (e.g., attention heads, MLPs) using causal mediation analysis techniques. Leveraging the extracted circuits, we first understand the interplay between the model's usage of parametric memory and retrieved context towards a better mechanistic understanding of context-augmented language models. We then identify a small set of attention heads in our circuit which performs reliable data attribution by default, thereby obtaining attribution for free in just the model's forward pass. Using this insight, we then introduce ATTNATTRIB, a fast data attribution algorithm which obtains state-of-the-art attribution results across various extractive QA benchmarks. Finally, we show the possibility to steer the language model towards answering from the context, instead of the parametric memory by using the attribution from ATTNATTRIB as an additional signal during the forward pass. Beyond mechanistic understanding, our paper provides tangible applications of circuits in the form of reliable data attribution and model steering.

Graph machine learning has witnessed rapid growth, driving advancements across diverse domains. However, the in-distribution assumption, where training and testing data share the same distribution, often breaks in real-world scenarios, leading to degraded model performance under distribution shifts. This challenge has catalyzed interest in graph out-of-distribution (GOOD) detection, which focuses on identifying graph data that deviates from the distribution seen during training, thereby enhancing model robustness. In this paper, we provide a rigorous definition of GOOD detection and systematically categorize existing methods into four types: enhancement-based, reconstruction-based, information propagation-based, and classification-based approaches. We analyze the principles and mechanisms of each approach and clarify the distinctions between GOOD detection and related fields, such as graph anomaly detection, outlier detection, and GOOD generalization. Beyond methodology, we discuss practical applications and theoretical foundations, highlighting the unique challenges posed by graph data. Finally, we discuss the primary challenges and propose future directions to advance this emerging field. The repository of this survey is available at https://github.com/ca1man-2022/Awesome-GOOD-Detection.

Emergent Communication (EC) provides a unique window into the language systems that emerge autonomously when agents are trained to jointly achieve shared goals. However, it is difficult to interpret EC and evaluate its relationship with natural languages (NL). This study employs unsupervised neural machine translation (UNMT) techniques to decipher ECs formed during referential games with varying task complexities, influenced by the semantic diversity of the environment. Our findings demonstrate UNMT's potential to translate EC, illustrating that task complexity characterized by semantic diversity enhances EC translatability, while higher task complexity with constrained semantic variability exhibits pragmatic EC, which, although challenging to interpret, remains suitable for translation. This research marks the first attempt, to our knowledge, to translate EC without the aid of parallel data.

Attribution tags form the foundation of modern cryptoasset forensics. However, inconsistent or incorrect tags can mislead investigations and even result in false accusations. To address this issue, we propose a novel computational method based on Large Language Models (LLMs) to link attribution tags with well-defined knowledge graph concepts. We implemented this method in an end-to-end pipeline and conducted experiments showing that our approach outperforms baseline methods by up to 37.4% in F1-score across three publicly available attribution tag datasets. By integrating concept filtering and blocking procedures, we generate candidate sets containing five knowledge graph entities, achieving a recall of 93% without the need for labeled data. Additionally, we demonstrate that local LLM models can achieve F1-scores of 90%, comparable to remote models which achieve 94%. We also analyze the cost-performance trade-offs of various LLMs and prompt templates, showing that selecting the most cost-effective configuration can reduce costs by 90%, with only a 1% decrease in performance. Our method not only enhances attribution tag quality but also serves as a blueprint for fostering more reliable forensic evidence.

Classification is a ubiquitous and fundamental problem in artificial intelligence and machine learning, with extensive efforts dedicated to developing more powerful classifiers and larger datasets. However, the classification task is ultimately constrained by the intrinsic properties of datasets, independently of computational power or model complexity. In this work, we introduce a formal entropy-based measure of classificability, which quantifies the inherent difficulty of a classification problem by assessing the uncertainty in class assignments given feature representations. This measure captures the degree of class overlap and aligns with human intuition, serving as an upper bound on classification performance for classification problems. Our results establish a theoretical limit beyond which no classifier can improve the classification accuracy, regardless of the architecture or amount of data, in a given problem. Our approach provides a principled framework for understanding when classification is inherently fallible and fundamentally ambiguous.

CapyMOA is an open-source library designed for efficient machine learning on streaming data. It provides a structured framework for real-time learning and evaluation, featuring a flexible data representation. CapyMOA includes an extensible architecture that allows integration with external frameworks such as MOA and PyTorch, facilitating hybrid learning approaches that combine traditional online algorithms with deep learning techniques. By emphasizing adaptability, scalability, and usability, CapyMOA allows researchers and practitioners to tackle dynamic learning challenges across various domains.

Contrastive learning has proven to be highly effective in pre-training large-scale models, especially in large vision models exemplified by frameworks like SimCLR [1, 2], MoCo series [3, 4], DINO series [5, 6]. However, with the ongoing scaling of large models, data hunger for contrastive learning is raising more attention in the community towards pre-training effectively from drift data. It could be caused by long-tailed data, noise, and domain shift, where concept drift [7, 8] is utilized to uniformly summarize this phenomenon of unpredictable distribution changes in the pre-training through contrastive learning. Hence, a pertinent question emerges: beyond the existing contrastive learning methods, can contrastive paradigm learn from drift pre-training? In this work, we aim to bridge this gap by providing a systematic analysis of the above question. Our findings highlight critical vulnerabilities of the current contrastive pre-training paradigm in adapting to these challenges, underscoring the need for novel strategies to enhance their robustness in drift data streams. More related works are provided in Appendix A. Current contrastive pre-training methods predominantly adhere to the paradigm of comparing two distinct views of the same object, typically derived from different encoders.